CN109271683A - A kind of groups of building automatic placement algorithm towards sunshine constraint - Google Patents
A kind of groups of building automatic placement algorithm towards sunshine constraint Download PDFInfo
- Publication number
- CN109271683A CN109271683A CN201810995301.0A CN201810995301A CN109271683A CN 109271683 A CN109271683 A CN 109271683A CN 201810995301 A CN201810995301 A CN 201810995301A CN 109271683 A CN109271683 A CN 109271683A
- Authority
- CN
- China
- Prior art keywords
- sunshine
- building
- environment
- constraint
- measuring point
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Geometry (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Architecture (AREA)
- Civil Engineering (AREA)
- Structural Engineering (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A kind of groups of building automatic placement algorithm towards sunshine constraint of the present invention, when planning is carried out in high density settling pond environment for one kind with architectural design, the automation algorithm of reply " how arranging multiple building concentrations (based on high level) to meet itself and the constraint of neighboring buildings sunshine " problem.Specifically: professional knowledge is applied according to arrangement, in the form of mathematical model, design can flexibly describe solution " environment " model of various arrangement problems, main body by single building as mobile decision, by the arrangement problem of all buildings with substep serial manner, it is described as to the mobile decision problem built one by one；Meanwhile design can realize " agent " model of mobile decision to single building therein with " environment " model by information interaction；Learn class algorithm using existing deeply, constructs " substep bout formula " that one " agent " interacts with " environment " and solve circulation.
Description
Technical field
When the present invention proposes that one kind carries out planning with architectural design in high density settling pond environment, how reply " arranges more
A building concentration (based on high level) is to meet the constraint of itself and neighboring buildings sunshine " automation algorithm of problem.
Background technique
With the continuous propulsion of Urbanization in China, high density settling pond space has become a kind of normality.In order to ensure
Vast civic healthy living, two-tier government of central and local governments all promulgates about ensuring different types of building, in 1 year
Specific date enjoys the designing technique specification of the minimum hourage of sunshine.For example, national regulation " norm of residential housing (GB 50368-
2005) have in " " standard that 4.1.1 Elder Housing is not lower than winter solstice sunshine 2h ", Shanghai place " Residence Design standard
(DGJ08-20-2007) " have that " small set, the middle set of 3.2.1 high residential building should at least have a living space, and big set is at least answered in
The winter solstice, which can be obtained, continuously there are two living space expires window effective sunshine not less than 1h."
For a long time, above-mentioned requirements, city planning design and designer are faced, is required at it in base range
Building arrangement during, be repeated using " trial-and-error method " " fine tuning arrangement with sunshine calculating ".Although between tier building
It only need to can disposably complete to arrange according to " the minimum distance for sunlight coefficient " in regulation, but be in built-up high density settling pond
In, it is in the high-rise problem in face of that can not arrange with the coefficient in most cases.On the one hand, inside and outside base
The shade of other skyscrapers can all influence the sunshine condition built in base；On the other hand, the skyscraper in base
It needs to consider simultaneously, to the solar radiation of buildings other inside and outside base.As it can be seen that high level during this arrangement, in base
Building positioning is the key that problem and difficult point.Since neighboring buildings situation all difference, shaped base that every secondary design is faced are made
Situations such as asd number again different at the spatial dimension that can arrange, that include in design big low height, is also different --- it is difficult
It is solved by establishing " white box " algorithm an of clear logic.So so far, still by time-consuming and laborious in design industry
Manual type completes above-mentioned arrangement task.
Therefore, in currently practical building arrangement, sunshine constraint is since it needs repeatedly the complexity of " trial and error ", relatively
For other design constraints (such as building spacing moves back boundary) without operating repeatedly, it has also become in high density settling pond construction, system
About design efficiency an important factor for one of.
For this problem, the national conditions faced by design abroad industry (land ownership system, environmental density and height
The design requirement of degree, all kinds of buildings) it is different from China, so without the directly available skyscraper towards sunshine constraint
Group's automatic placement algorithm.And though there are some relevant research achievements for the country.
One kind is that can automatically generate building arrangement as a result, but being not related to the correlative study that sunshine constrains.
For example, Li Biao of Southeast China University et al. has been carried out building formation and has been set based on " class rectangle base subdivision " method
Meter, any given building landform is subdivided into it is multiple approximate rectangular, then wherein in a self-adaptive manner insertion have China pass
The groups of building for local-style dwelling houses feature of uniting generate the building block for covering whole bases[1], focus is more body itself, and
Non- sunshine constraint.
For another example, Tongji University Sun Cheng space et al. is arranged by the shape and building of analyzing a large amount of urban residential land cases
Cloth type parameterizes prototype by foundation " neighbour, form a team, build " three-level, realize do not consider under sunshine constraint condition from opening
The automatic placement that mark of bristling with anger arrives three-dimensional building mathematical model generates.[2]
It is another kind of be can automatically generate built arrangement by sunshine constraint as a result, but the arrangement problem that can handle have
Apparent limitation can not adapt to actual design requirements of one's work.
For example, Song little Dong of Tongji University et al. in 2010, attempts to carry out computer under the constraint of existing sunlight ordinance
Automatically generate residential housing layout.It has used for reference " two-dimentional bin packing algorithm " to the rectangle model for including building with " spacing control zone "
Enclose carry out automatic placement, and place one's entire reliance upon for the satisfaction of sunshine constraint condition " spacing control zone " magnitude estimation on.It is aobvious
So, due in algorithm " spacing control zone " be to be estimated according to the outsourcing rectangular extent of building concentration profile, empirical parameter, from
The sunshine shade superimposed factor with neighboring buildings can not be fundamentally taken into account, so the region can only be amplified, moderately just also with true
Sunshine constraint can be met in greater probability by protecting generation result --- and this directly contributes the automatic placement algorithm can not really sufficiently
The advantage being overlapped mutually using sunshine shade improves arrangement density, to adapt to the actual design demand of high density settling pond.[3]
For another example, Ling Ling of the Central China University of Science and Technology et al. 2013 according to settlement building sunshine in need of consideration, disappear
Anti- equal constraint, constructs the mathematical model of architectural composition, realizes determinant and the automatic fast layout of perimetric pattern.It, should according to introduction
The processing that method constrains sunshine, which is only embodied in, to be taken " minimum distance for sunlight coefficient 1.2 " disposably to control between building concentration
Positional relationship, and the practical shade superposition arrangement problem of high level above-mentioned can not be coped at all, actual design can not be applied to.[4]
Also for example, " forced-ventilated " technology of the publication of Shenzhen little Ku science and technology in 2017 uses " volume according to its publicity literature introduction
Product neural network (Convolutional Neural Network, CNN) " " generates confrontation network (Generative
Adversary Network, GAN) " two kinds of deep neural network technologies.It is instructed by arranging case with a large amount of outstanding buildings
Practice neural network, so as to then according to a given base condition, automatically generating out a batch arrangement as a result, wherein
A part in high probability (i.e. its so-called " certified product rate "), every design requirement can be met, here just include day
According to constraint condition.[5] obviously it is inferred that since the arrangement result that can be automatically generated here places one's entire reliance upon for trained
Existing case characteristic, that is, require that (including place latitude, boundary shape, plot ratio, building are close for various potential base conditions
Degree, building concentration quantity, chamfered shape, height, periphery existed building condition etc.) it requires to find a large amount of corresponding training case
Example --- this be under the conditions of changeable actual design it is unpractical, i.e., there are inadequate natural endowments in adaptability for this method.
Prior art index:
[1] technology of Li Biao, Han Dongqing building formation design understands and its prospect [J] building journal, and 2011,06:96-
100.
[2] Sun Chengyu, Luo Qiming, Song little Dong, Xie Junmin, the city threedimensional model side of automatically generating for mirror face the practice of having mercy on
Method --- the by taking the intensity block planning of Beihai as an example builds journal, 2017,8:77-81.
[3] Song little Dong, Pang Lei, Sun Cheng space house plot plot ratio evaluation method visit [J] urban planning academic periodical again,
2010,02:57-63.
[4] Ling Ling, Wang Xiaobo, Lee are calculated at rigid based on settlement building automatic layout algorithm [J] of internal feasible zone
Machine technology and development, 2013,08:1-4.
Does [5] small library science and technology .AI change century-old automobile industry? Millennium building industry is also required to rescue [EB/OL] .:http: //
mp.weixin.qq.com/s/8XKKyxjgBU8AkZkkD5iMQA,2017.12.27:
[6]Timothy P.Lillicrap,Jonathan J.Hunt,Alexander Pritzel,Nicolas
Heess,Tom Erez,Yuval Tassa,David Silver,Daan Wierstra.Continuous Control with
Deep Reinforcement Learning[EB/OL].https://arxiv.org/abs/1509.02971,
2016.2.29:
Summary of the invention
The present invention needs technical solution characterization to be protected are as follows:
Technical solution one
A kind of building body automatic placement design method that can be constrained towards computability, which is characterized in that answered according to arrangement
With professional knowledge, in the form of mathematical model, design can flexibly describe solution " environment " model of various arrangement problems, will be single
A building is as the main body for moving decision, by the arrangement problem of all buildings with substep serial manner, is described as to building one by one
The mobile decision problem built；Meanwhile design can pass through information interaction with " environment " model, and to single building therein
Realize " agent " model of mobile decision；Learn class algorithm using existing deeply, constructs one " agent " and " ring
" the substep bout formula " of border " interaction solves circulation: when each bout starts, all buildings in " environment " are in an initial row
Cloth state, and in subsequent each step, " agent " basis is interacted with " environment ", carries out the shifting of each building one by one
It is dynamic；In arbitrary steps, the current state of " agent " acquisition " environment ", to generate the single step mobile vector individually built,
After executing the movement in " environment ", then the new state after movement is obtained, and mobile front and back " environment " is in a certain computability
The variable quantity of satisfaction degree on capable of constraining.The moving recording updated with above-mentioned dynamic (includes at least: before mobile vector, movement
The variable quantity of " environment " state, mobile front and back performance constraints satisfaction degree afterwards), to the depth nerve net in mobile decision function
Network is trained, and to obtain higher performance constraints satisfaction degree in the bout later period, i.e., is mentioned from the mobile experience of Dynamic Accumulation
Rise the ability for generating correct mobile vector.With the increase of rounds and number of steps, this performance indicator will be continuously improved, until
The solution of arrangement problem is completed, that is, reach this performance constraints being previously set meets condition.In above-mentioned performance constraints
Performance indicator, it is necessary to be the numerical value class index that acquisition can be calculated from " environment ".
Technical solution two
A kind of groups of building automatic placement algorithm towards sunshine constraint, which is characterized in that
According to the professional knowledge in building sunshine field, " the ring that can flexibly describe various groups of building arrangement problems is devised
Border " mathematical model (A environment) devises " agent " control mould that can obtain the interaction experience of the single stage at this in " environment "
Block (B agent) constructs one based on " agent " and " ring to can learn class algorithm according to existing deeply
" substep bout formula " the deeply learning training that border " interacts experience solves module (C substep bout formula).Wherein, " agent "
Constantly after each " single stage " the middle state for obtaining " environment ", the mobile decision function that input control is individually built, it is driven
It is mobile, and the global sunshine constraint satisfaction degree variable quantity of mobile front and back, and new " environment " state are collected, by these
The movement decision function is continued to optimize in the record of interaction experience and training, to reach global day in subsequent bout and step
According to the top of constraint satisfaction, until completing the solution of sunshine arrangement problem.
Above technical scheme is completely independent of the arrangement sample collected in advance, has to all kinds of arrangement tasks high
Adaptability, the byproduct (the mobile decision neural network of single moving body) solved have the very high transfer of learning.
Detailed description of the invention
Fig. 1 algorithm flow chart
Fig. 2 environmental model sample
Fig. 3 sample groups of building hiding relation general layout on sunshine measurement direction at a certain moment
Hiding relation digraph on Fig. 4 sunshine measurement direction
Fig. 5 determines the neighbouring relations of figure interior joint with maximum hiding relation
N+m positive and negative measurement directions (n=9, m=3) of the Fig. 6 about practical movable distance around
Specific embodiment
Technical solution of the present invention is further described below in conjunction with attached drawing.
Existing skyscraper position automatic placement method and correlative study towards sunshine constraint, is all unable to satisfy reality
The adaptability requirement that design work proposes arrangement task.Technical solution of the present invention:
For this purpose, relevant professional knowledge of the present invention according to building sunshine field, various buildings can flexibly be described by devising
" environment " mathematical model (A frame portion point in Fig. 1) of group's arrangement problem, it is mutual to devise the single stage that can be obtained at this in " environment "
" agent " control module (B frame portion point in Fig. 1) of dynamic experience, so that class algorithm can be learnt according to existing deeply
(by taking DDPG algorithm as an example) constructs " the substep bout formula " deeply for interacting experience with " environment " based on " agent "
Learning training solves module (C frame portion point in Fig. 1).Wherein, " agent " constantly obtains " environment " in each " single stage "
State after the mobile decision function that input control is individually built, drives its movement, and collects the global sunshine constraint of mobile front and back
Satisfaction degree variable quantity continues to optimize the movement decision function, thus rear by the way that these are interacted with record and the training of experiences
Reach the top of global sunshine constraint satisfaction in continuous bout and step, until completing the solution of sunshine arrangement problem.The calculation
Method is completely independent of the groups of building arrangement case sample collected in advance, has high adaptability to all kinds of arrangement tasks,
The byproduct (the mobile decision function individually built, i.e. neural network) of solution also has very high reusability, that is, transfer of learning
Property.
So-called " agent " of the invention is known as " Agent " in the deeply theories of learning.
So-called " adaptability " of the invention, is embodied in:
It can either be according to mission requirements, in the given position setting sunshine measuring point of each building concentration for needing to arrange and day
According to constraint condition --- adapt to any computational accuracy；
Various arrangement problems are coped with again, and being not limited to latitude, boundary shape, topography, just variation, plot ratio, building are close
Degree, building concentration quantity, building concentration type, chamfered shape, height, periphery existed building condition etc. --- adapt to various arrangements
Condition.
Using: algorithmic technique scheme through the invention can support corresponding computer software, to carry out high density
The designer of urban planning or architectural design, to given building lot range and multiple skyscraper monomers and it is a series of its
Its design requirement, carries out the automatic placement schemes generation for meeting sunshine constraint condition, and designer can be from inefficient repeatedly artificial
It is freed in " trial and error " operation, suitable arrangement is selected in the result automatically generated, as subsequent Deepen Design
Basis.
The present invention provides specific technical solution
The present invention can be summarized as one and be interacted by " the substep bout formula " of A " environment ", B " agent ", C between the two, and three
The circulation solution procedure (Fig. 1) being grouped as.
A " environment ": " environment " mathematical model of description groups of building arrangement problem.
By any one " the groups of building automatic placement towards sunshine constraint " problem, it is described as one based on three-dimensional system of coordinate
Three-dimensional " environment " mathematical model of (such as X-axis is directed toward positive east, and Y-axis is directed toward the positive north, and Z axis points into the sky) is (afterwards referred to as " environment ",
Content in left-hand broken line frame in Fig. 1), it is partially used in solution procedure for subsequent B " agent ".It includes following 5 kinds of objects:
1) sunshine measuring point: one group of three-dimensional point { P1, P2, P3 ... }, the form to discrete three-dimensional point describe those needs
The area of space for considering sunshine constraint has the black of digital number such as the central point of building two windows of bottom each in Fig. 2
Solid circles.
2) sunshine measurement direction: one group of n trivector describes continuous sunshine process in one day to discrete form.
Each vector corresponded to one it is specific at the time of, be directed toward the direction of the sun (by place city latitude from sunshine measuring point
Influence), as in Fig. 2 from 9 vectors issuing southwards of building No. 0 window center point of E bottom, they respectively represented from 8:00 to
The solar vector (by taking the latitude of Shanghai as an example) at 16:00 integral point moment.
3) sunshine constraint condition parameter: one group of numerical parameter associated with each sunshine measuring point and logic judgment rule,
The different sunshine constraint condition of each city, all types of buildings is subjected to mathematical description, comprising:
Correspondence time interval between adjacent sunshine measurement direction: numerical value shape parameter, it describes the adjacent sunshine of any two
The interval of time representated by measurement direction.As the value reduces, discrete sunshine measurement direction becomes more intensive, the day of this algorithm
Higher according to simulation precision, calculation amount also increases.
The constraint satisfaction Rule of judgment of sunshine measuring point: one group of logic judgment rule, it is directed to different cities, sunshine measuring point institute
In different sunshine constraint conditions of the building type corresponding to it, describe whether sunshine constraint meets in the form of rules.Including most
The satisfaction at early effective sunshine moment judges that the satisfaction judgement at effective sunshine moment, accumulative effective sunshine minimum duration expire the latest
Foot judgement, the satisfaction judgement of continuous effective sunshine minimum duration.
4) building: one has the geometry polyhedron of multinomial attribute data, and sunshine can be blocked in environment to describe
Building entity part (five of A to the E rectangles for having oblique line texture are indicated in Fig. 2).Its attribute includes:
High-rise attribute: yes/no assigns automatically according to polyhedral Z axis size)；
Removable attribute: yes/no, when being only, which just participates in automatic placement calculating；
Sunshine measuring point list attribute: numerical value array, number of the description relevant date according to measuring point；
Minimum level boundary: a closed curve in X/Y plane, description need full when two buildings move horizontally
The minimum boundary that foot is not overlapped mutually requires (respectively rectangular extent of the building with oblique line rectangle north side in Fig. 2).
5) neighbour boundary: 4 closed curves in X/Y plane describe the boundary of neighbour and road, multilayer in neighbour respectively
It builds the range boundary that can be arranged, the range boundary that skyscraper can arrange in neighbour, allow range of arranging in neighbour
Boundary (such as concentration greenery patches, water body), such as 1 solid line and 3 irregular ranges of dotted line in Fig. 2 from outside to inside.
B " agent ": " agent " control module of the single stage interaction experience in " environment " is obtained
By with the interacting of above-mentioned " environment " (read or modification model in data), complete in a single stage, it is right
The control (B1-B11 is walked in dashed centre line frame in Fig. 1) of any one building to be arranged, includes four class functions: calculating global day
According to constraint condition satisfaction degree (B1, B11 step in Fig. 1), selection currently building (B2 is walked in Fig. 1) to be moved, calculating to work as
" environment " state (B3-B9, B11 are walked in Fig. 1) obtained is observed centered on preceding building, is executed and is currently built in " environment "
Movement under given vector (B10 is walked in Fig. 1).
" the substep bout formula " interaction of C between the two: the deeply learning training based on the two interaction experience solves module
" substep bout formula " general solution frame common in deeply study is used for reference (in Fig. 1 in the C dotted line frame of right side
C1-C10 step, by taking DDPG algorithm as an example [6]): there is u " bout ", often between building one " environment " and " agent "
The interactive process of bout v " step ".Period, each " bout " are restored to the original state of " environment " when starting, and in office
In one step t, " agent " is accumulated in " environment " in relation to specific behavior condition, behaviour decision making, the interaction experience of Behavioral effect
(i.e. specific behavior executes " environment " state tensor St and St+1 of front and back, is generated by the behavioral decision function of " agent " sample
The control tensor A t of specific behavior, " agent " behavior execute the specific benefit variable quantity Rt+1-Rt of front and back " environment ").When
After interaction experience sample size reaches requirement, in each step, so that it may be used to in the behavioral decision function of " agent "
Deep neural network be trained, i.e., in interactive process, the behavior of one " agent " in " environment " is presented as, from nothing
Purpose is explored to the continuous learning process for obtaining higher specific benefit.Then, after several " bouts " several " steps ", when specific
When benefit reaches solution and requires, i.e. declaration problem solving is completed.
By the General Principle of above-mentioned deeply learning algorithm, constructed with above-mentioned according to building sunshine domain-specific knowledge
Specific " environment " to get up combines that (i.e. the state tensor of " environment " specifically, be with selected single build with " agent "
Center, the neighboring buildings circumstance of occlusion observed and each description to potential movable distance；The specific behavior control of " agent "
Tensor processed is specifically, selected individually builds the single step mobile vector in mobile range；The specific benefit of " environment " changes
Amount is specifically, the mobile front and back of selected single building, the variable quantity of global sunshine constraint condition satisfaction degree), it is formed this
" the substep bout formula " solution procedure of algorithm for " sunshine constrains lower groups of building automatic placement " problem.
Further provide detailed embodiment algorithm
Embodiment 1
Thought based on technical solution of the present invention provides the embodiment algorithm for using DDPG to learn for deeply, but not
The scope of the present invention should be limited with this.
This algorithm converts the Parallel implementation process between buildings multiple in automatic placement problem to for single building movement
The deeply learning process of decision function.The essence of entire learning process, is exactly the initial arrangement constantly from " bout ",
Single building movement in a series of " steps " in " agent " control " environment ", records mobile behavior in each step
The variation of amount, the different conditions of mobile front and back " environment ", overall situation sunshine constraint condition satisfaction degree, is instructed with these sample datas
The mobile decision function for practicing " agent " is restrained towards the higher global sunshine constraint satisfaction degree of acquisition, thus multiple
After " bout " and " step ", realizes and solve.Specifically, solution procedure includes to have v " steps in u " bout " and every bout
Suddenly nested circulation ", i.e., have (Fig. 1) for any t step (e ∈ [0, u-1], t ∈ [0, v-1]) in any e bout:
Step 1: (again) setting solves hyper parameter
According to available computational resources, various solution hyper parameters are set, such as: the rounds u of calculating process, every bout number of steps
V, the neuronal quantity in deep neural network framework, connection type, activation primitive type etc..
Step 2: judge the termination condition of bout circulation
If e=u, i.e. this bout the last one bout for having exceeded setting, and before also without occurring that solution can be used, then this
Wheel is solved with failure terminating.If (cooperation hyper parameter automatic search method, can return to step 1, new one is attempted after resetting
Wheel solves.)
Step 3: start a new bout k
The building in environment is subjected to initial layout by setting rule (as being randomized), and step counter t is reset；
Step 4: the termination condition of judgment step circulation
If t=v, i.e., this step has exceeded the last one step in the bout of setting, then bout counter e adds " 1 " (e
=e+1), and return to step 2；
Step 5: start the interaction step t of one new " agent " and " environment "
Step 6: global sunshine constraint satisfaction percentage Rt is calculated
By making ray along n different sunshine measurement directions respectively, judging it from each sunshine measuring point in environment
Whether in environment building formed geometrical intersection relationship, i.e., analog, which calculates all sunshine measuring points, can actually obtain day
According to details.Further according to sunshine constraint condition parameter, can count meet sunshine constraint sunshine measuring point quantity account for
The ratio of all sunshine measuring point quantity, i.e., global sunshine constraint satisfaction percentage Rt when step t starts.
Step 7: current building to be moved is selected
Strategically selected (as random) one removable building b is as the current wait move of this step in " environment "
Dynamic building.The calculating of " environment " state Sbt when all subsequent t to current procedures is all centered on it.
Step 8: global sunshine resource digraph is constructed
Construct n hiding relation digraph { G1, G2...Gn } of all sunshine measuring points on n sunshine measurement direction.Its
In, Subgraph Gn constructs that (any " environment " in District of Shanghai in such as Fig. 3, Fig. 4 are it when 13:00 by the following method
The digraph G of sunshine measurement direction): the floor projection vector of the sunshine measurement direction corresponding along figure traverses in scene and owns
Relationship (any " environment " in District of Shanghai, the sunshine test side when 13:00 in such as Fig. 3 of sunshine measuring point and neighboring buildings
To floor projection traversal).It is using sunshine measuring point as the node of figure, using sunshine hiding relation as the side of figure, with from being blocked
Sunshine measuring point is that (such as Fig. 4 is that having for " environment " is somebody's turn to do on above-mentioned direction to positive direction to the sunshine measuring point that the building for blocking it is possessed
To figure G).When in figure a measuring point node i and one, its day side building form maximum hiding relation (rise is than maximum),
Then (such as measuring point node j) forms neighbouring relations in figure and (is such as directed toward and saves from node i all measuring points on the measuring point and the building
(such as Fig. 4 is the A:1 node of above-mentioned figure to point j), although being blocked simultaneously by building B and C in southern side, because of the rise ratio with building C
It is bigger, so the positive direction node of the node is C:0 and C:1, rather than B:0 and B:1).In rise ratio herein, distance Dij
It is the ground length along path from building where the sunshine measuring point node i to adjacent node j being blocked, constituting hiding relation to it
Degree.Height difference Hij be from building where the sunshine measuring point node i to adjacent node j being blocked, constituting hiding relation to it,
The vertical range (such as Fig. 4) at the practical top point of contact under direction sun three-dimensional ray.
Step 9: day side sunshine measuring point node is extracted
In each digraph Gn, respectively with currently move building possessed c sunshine measuring point node P1,
P2...Pc } it is starting point, along its direction that is blocked (positive direction in figure, arrow direction) between Fig. 4 interior joint, find all positions
In its day side, other sunshine measuring point nodes in the building of direct or indirect hiding relation are constituted to it.
Step 10: it calculates day side " sunshine resources index "
For the above-mentioned node extracted from each digraph Gn, calculates the wherein sunshine corresponding to each node and survey
Point ground distance Di on corresponding sunshine measurement direction and blocks height between building corresponding to its day side adjacent node
Poor Hi, then day side " sunshine resources index (+Solar of the current building on current digraph Gn is calculated as follows out
Index ,+SI) " (formula 1).Finally, n+SI will be obtained for n sunshine measurement direction.
Step 11: it extracts and carries on the back positive side sunshine measuring point node
In each digraph Gn, respectively with currently move building possessed c sunshine measuring point node P1,
P2...Pc } it is starting point, the direction (opposite direction in figure) blocked inversely is formed along to other buildings, is found all positioned at its back
Positive side, by the sunshine measuring point node for other buildings that building where it is directly or indirectly blocked.
Step 12: it calculates and carries on the back positive side " sunshine resources index "
For the above-mentioned node extracted from each digraph Gn, calculates the wherein sunshine corresponding to each node and survey
Point ground distance Di on corresponding sunshine measurement direction and blocks height between building corresponding to its day side adjacent node
Poor Hi, then back sun side " sunshine resources index (- Solar of the current building on current digraph Gn is calculated as follows out
Index ,-SI) " (formula 2).Finally, n-SI will be obtained for n sunshine measurement direction.
Step 13: it calculates each to movable distance
From the centroid currently built, along n+m both forward and reverse directions, (n is sunshine measurement direction, such as the positive negative side of n=9 in Fig. 6
To long arrow；M is additional directions, such as the short arrow of the both forward and reverse directions of m=3 in Fig. 6) floor projection, calculate separately they arrive at
It (is calculated according to the corresponding neighbour boundary of the high-rise Attributions selection of building) in the distance on corresponding neighbour boundary, it actually can be with
Mobile distance (deducting by the occupied line segment overall length of the minimum level bounds of building therebetween).2n+2m is obtained in total
A distance value D.
Step 14: the state Sbt of synthesis " environment "
Above-mentioned all+SI ,-SI, D, constitute the array with 4n+2m element, i.e., with fixed quantity parameter
Form describes the global context state Sbt observed centered on currently building.
Step 15: action vector Abt is generated with decision function
After Sbt is inputted the adjustable mobile decision function Fae of an inner parameter, the vector Abt currently moved is calculated
Present valuation (formula 3).Here, as mobile decision function is during study (i.e. the training of deep neural network), inside
Parameter undergoes one from random distribution, and to gradually convergent process, calculated vector can also make the movement of current monolithic building,
From random trial, gradually tend to purposive exploration, and then obtains preferably global sunshine constraint satisfaction percentage；
Step 16: the movement currently built under given vector in " environment " is executed
Current building b is carried out by vector Abt present valuation in the environment practical mobile；
Step 17: the state Sbt+1 and sunshine constraint satisfaction percentage Rt+1 of " environment " after movement are calculated
According to the step identical as step 7-14, again centered on the building b after currently moving, neighboring buildings are observed,
The global state for calculating new position describes Sbt+1 (be equivalent to building b and enter step the global state observed when t+1)；Meanwhile
Calculate the percentage Rt+1 that this stylish all sunshine measuring point meets sunshine constraint；
Step 18: storage interaction experience
By above-mentioned data (Sbt, Abt present valuation, Sbt+1, Rt+1-Rt) deposit, one has the list M of fixed capacity (to ask
List is empty when solution starts, then constantly deposit；Most newly generated data replace wherein oldest data after being filled with)；
Step 19: interaction experience is trained and carrys out Optimal Decision-making function
If list M has expired, learning training is unfolded to mobile decision function.
The present embodiment is set according to DDPG algorithm: the control tensor (list i.e. here of the specific behavior of all " agents "
The mobile control vector of a building) by one group of adjustable deep neural network of inner parameter (mobile decision function) Fae, Fat,
Fce, Fct are calculated.Function Fae, Fat calculate the control tensor A (formula 3, formula 4) of behavior according to the state tensor S of " environment ".Letter
Number Fce, Fct are responsible for calculating the value (formula 5, formula 6) after (i.e. mobile) execution of behavior according to state tensor S, tensor A.Preceding two
Between a function, deep neural network structure having the same between latter two function, and each coefficient is carried out periodically by algorithm
It synchronizes (from Fae to Fat；From Fce to Fct).There are the relationship of formula 7, (GAMMA is between 0 to 1 between t step and t+1 step
The parameter of value, for determining the acquired return (Rt+1-Rt) of current procedures value relatively at a specified future date, the specific gravity in decision,
There are " foresight " [6] more tending to 1 study).Then, so that it may using FCEloss as the loss function (formula of function Fce
8), the loss function (formula 9) using FAEloss as function Fae, carries out the training of deep neural network, final to can get 4 certainly
The convergence state of plan function.
Abt present valuation=Fae (Sbt) (formula 3)
Abt+1 long-term goal=Fat (Sbt+1) (formula 4)
Qt present valuation=Fce (Sbt, Abt present valuation) (formula 5)
Qt+1 long-term goal=Fct (Sbt+1, Abt+1 long-term goal) (formula 6)
Qt long-term goal=Rt+1-Rt+Qt+1 long-term goal * GAMMA (formula 7)
FCEloss=| Qt long-term goal-Qt present valuation | (formula 8)
FAEloss=-Qt present valuation (formula 9)
Step 20: judge whether solution succeeds according to Rt+1
If judging that current Rt+1 has reached the requirement of Solve problems counterglow constraint, the current arrangement in environment is remembered
Record be successfully to solve, at this time " substep bout formula " solution procedure can successful termination (if necessary to it is multiple meet sunshine constrain solutions, then
Can be after record arrangement, termination algorithm, does not continue to execute subsequent 21st step).
21st step: enter next step
Step counter t adds " 1 " (t=t+1), and returns to step 4.
Key point of the invention
1) " environment " mathematical model for describing arrangement problem, can flexibly express various arrangement problems.
In the A " environment " of algorithm, the three-dimensional mathematical model formed using 5 kinds of objects can pass through the difference of its each parameter
Combination settings flexibly express the various groups of building arrangement problem under the various sunshines constraint in actual design.
2) more buildings that arrangement problem is defaulted are described to be changed to the serial description of single building parallel, reduces and solves difficulty.
The C " environment " of algorithm is interacted with " the substep bout formula " between " agent ", by building arrangement task, is generally defaulted
The parallel orientation problems of disposable multiple buildings, be converted into random sequence, the single building of " substep bout formula " moves
Decision problem.According to the analysis above to domestic and international correlative study, existing method all attempts directly to produce one comprising all
The arrangement result of building location information in base range to be arranged.It is limited by non-linear space distribution, the difference of sunshine condition
Common Complex Constraints condition in practical coincidence judgement and various actual designs between the building of base profile, so that this is asked
The difficulty of topic remains high.And the present invention is asked by building one " the substep bout formula " with single building for mobile decision package
Solution preocess converts the above problem to, and any single building is taken appropriate mobile tensor, gradually changed according to surrounding enviroment state
Kind overall situation sunshine condition, it is single to build the mobile decision problem of single step.Thus greatly reduce the grade of difficulty of problem.
3) " sunshine resources index " is used to describe the global sunshine on 1 set sunshine measurement direction with 2 numerical value and hide
Keep off comprehensive condition.
To step 12, the neighboring buildings in face of the day side of each sunshine measurement direction and the positive side of back become step 8 in algorithm
It is dynamic, series of steps is proposed, to use the digraph based on sunshine measuring point, cumulative calculation wherein builds associated section with current
The day side " sunshine resources index " of point and the positive side " sunshine resources index " of back.It observes " environment " for current building, provides
The ability that global sunshine on 1 direction blocks comprehensive condition, i.e. n sunshine measurement direction are described with fixed quantity (2) numerical value
And under normal circumstances, (on a sunshine measurement direction, quantity and position are all for the ability described with fixed 2n numerical value
Changed neighboring buildings or the numerical value for using fixed quantity, to describe the Partial Perimeter currently built, in fixed quantity
Discrete direction on, the case where being blocked by sunshine, i.e., the overall situation can not be described；Same variable number numerical value is used, is come
Brought sunshine circumstance of occlusion is dynamically built on description periphery, i.e., can not describe the overall situation with regular length numerical value).Due to this
There are high correlations between the variation for the solar radiation for bringing neighboring buildings after information and current building are mobile, so it is for each
" mobile decision function " study of kind of deeply learning algorithm can all be played the role of key, be independently of certain depth reinforcing
The professional description innovation of the building field of learning algorithm.Meanwhile it contains " ring as a kind of statistical value from rise ratio
The elevation information in border " has the ability for treating different height building with a certain discrimination, has processing and asks with the arrangement of topography height difference
The ability of topic.
4) global context state is described using regular length array.
Step 7 and step 14 in algorithm use regular length (comprising 4n+2m member in face of different arrangement problems
Element, 2n description " sunshine resources index ", 2n+2m describes that the range of distance can be moved) array, to be retouched in discrete form
It states centered on currently building, the global context observed.This is the net between similar arrangement task, in migration " decision function "
Network training result establishes basis.If (data length of ambient condition description changes because task of arranging is different, can not just allow
It shares a success between problem is similar but surrounding enviroment are different " decision function " neural network trained.)
5) balance between computational accuracy and time-consuming is adjustable.
The present invention, can be by adjusting the sunshine measuring point quantity in environment, sunshine test side for set arrangement task
Rounds u and number of steps v into quantity n, deeply learning framework, to change the precision of solution, and correspond
Calculating it is time-consuming, adapt to each design phase for the different demands of precision and speed.
6) the solution acceleration mechanism under condition of similarity.
Deep neural network structure is used in " mobile decision function " of the invention, has benefited from its training result and generally has
" study transitivity " characteristic having, and single global state description for building heuristic process, regular length of " substep bout formula ",
It makes it possible in the case where two similar arrangement task (sunshine measurement direction are identical as sunshine constraint condition parameter), directly
The deep neural network that first tasking learning is come, copies to the deep neural network of second task, to significantly reduce
The calculating time of second arrangement task solving.
The present invention has what advantage or beneficial effect
1) independent of sample.
The present invention (relies on the accumulation and training of high-quality sample, it is intended to disposable to generate using the prior art is totally different from
The successful arrangement result of multiple buildings) thinking, solution procedure can be described as: a building by ANN Control
It arranges " agent ", its rule is passed in condition characterising parameter only in accordance with the mankind, by a 3-d mathematics " ring
Substep bout trial is constantly carried out in the model of border ", the strategy individually built preferably is moved in acquistion from the experience of oneself, finally
Complete the process of entire arrangement task.And the present invention does not need thoroughly artificially to provide sample, it also would not be by artificial sample product
The brings such as tired, type missing restrict.Theoretically, the arrangement problem that the present invention is capable of handling will not be weak in complexity
In the mankind.
2) better adaptability.
It is unit that the object of prior art study, which is neighbour, and the result of generation is also a complete neighbour, and actual design
In arrangement problem need to consider many existing, set buildings inside and outside base and the various topographies inside and outside base toward contact
Height is distributed, so prior art day, which is born with, to be handled.And the present invention is various to describe by " environment " mathematical model
Arrangement problem, the training learning process of solution are the mobile decision letter of the single individually built for customizable sunshine measuring point again
Number, so not restricted by the type, form, quantity of building to be arranged completely and (only relating to solve the variation for calculating the time).Than
Such as, it for different latitude and local regulation requires different cities, for example is arranged in the opening in high-rise residential buildings of Shanghai City and Changchun
In cloth, " 1 hour sunshine between winter solstice 9:00-15:00 ", " 2 small time between Great Cold day 8:00-16:00 are just required respectively
According to ", it can be adapted to completely by the sunshine constraint condition parameter being respectively set in above-mentioned " environment "；It works for actual design
In face various shaped bases, arrangement asd number, building type, can by setting environment in building and attribute,
The objects such as neighbour boundary adapt to；It, can be by sunny slope when not yet determining window position in future architecture plane when for arrangement
Uniformly distributed sunshine measuring point, and when there is determining architectural plane, then sunshine measuring point directly can be set by window position；For some bases
(either it can cause that sunshine is blocked or its own sunshine will receive it to other buildings for existing intended object inside and outside ground
Its Influence on Architecture needs to be protected), can be by the building in setting environment, and corresponding attribute is set to realize.
Meanwhile the prior art is limited by the description precision of training sample, the precision of the result of generation can not necessarily surmount.And
The present invention for set arrangement task, can by adjust the sunshine measuring point quantity in environment, sunshine measurement direction quantity n,
Rounds u and number of steps v in deeply learning framework, to change the precision and corresponding calculating consumption of solution
When, each design phase is adapted to for the different demands of precision and speed.For example, between adjacent sunshine measurement direction in environment
Corresponding time interval can be used to control computational accuracy (interval is smaller, it is longer to calculate time-consuming, but solving precision is then higher),
Design early stage uses biggish value, carries out the rough quick arrangement of more wheels, and stage uses lesser value, with longer calculating
Time obtains more accurate arrangement result.
3) looser training result is multiplexed condition.
The prior art is limited to its generation confrontation network mechanism used, it depends on training result type and the place of sample
Generation when managing arrangement problem solves the necessary matched of type, otherwise just generates useless solution --- the multiplexing condition of training result
It is more harsh, for example, small-scale base arrangement training result can not be applied to extensive base arrangement task, L-type base
Training result of arranging can not be applied to the arrangement task in trapezoidal base.And the present invention uses method that single building moves one by one,
The global context state description of regular length, so that if joining between two arrangement tasks in sunshine constraint condition having the same
It counts, when being in the same city, and being equipped with identical computational accuracy (identical n sunshine measurement direction), so that it may be instructed
Practice achievement multiplexing --- multiplexing condition is looser, and " transfer of learning " for being conducive to give full play to neural metwork training achievement is excellent
Gesture greatly reduces the calculating time that second task generates solution for the first time.
The realization of this algorithm (can be not limited to Actor Critic, A3C, Deep based on various deeply learning algorithms
Deterministic Policy Gradient, that is, DDPG, Deterministic Policy Gradient, that is, DPG,
Proximal Policy Optimization, that is, PPO, Distributed Proximal Policy Optimization are
DPPO,Policy Gradient,Normalized Advantage Functions；In addition, when being in what " agent " generated
Continuously distributed mobile vector after being changed to multiple discrete fixation mobile vectors, can also use the calculation of more early stages a series of
Method, such as Deep Q Network, that is, DQN, Dueling DQN, Double DQN, Prioritized replay DQN).Obviously,
After specifically selecting different deeply learning algorithms, convergence speed and the success rate of neural network be will be different,
Then the solution efficiency of arrangement is influenced.Since the solution frame (step in the C portion dotted rectangle of right side in Fig. 1) of algorithm follows
The General Principle of deeply learning framework, and it is not the key point and protection point that the present invention advocates, so if using
Other deeply learning algorithms are trained " the mobile decision function " individually built, and should also regard that one kind of the invention is prolonged as
It is continuous.
Claims (11)
1. a kind of building body automatic placement method that can be constrained towards computability, which is characterized in that according to arrangement application profession
Knowledge, in the form of mathematical model, design can flexibly describe solution " environment " model of various arrangement problems, will individually build
Main body as mobile decision is described as the shifting to building one by one by the arrangement problem of all buildings with substep serial manner
Dynamic decision problem；Meanwhile design can realize single building therein and move with " environment " model by information interaction
" agent " model of dynamic decision；Learn class algorithm using existing deeply, constructs one " agent " and interacted with " environment "
" substep bout formula " solve circulation: when each bout starts, all buildings in " environment " are in an initial arrangement state,
And in subsequent each step, " agent " basis is interacted with " environment ", carries out the movement of each building one by one；Any
In step, the current state of " agent " acquisition " environment ", to generate the single step mobile vector individually built, in " environment "
After executing the movement, then the new state after movement is obtained, and mobile front and back " environment " is on a certain computability can constrain
The variable quantity of satisfaction degree；The moving recording updated with above-mentioned dynamic carries out the deep neural network in mobile decision function
Training is promoted that is, from the mobile experience of Dynamic Accumulation and is generated to obtain higher performance constraints satisfaction degree in the bout later period
The ability of correct mobile vector；With the increase of rounds and number of steps, this performance indicator will be continuously improved, until the row of completion
The solution of cloth problem, that is, reach this performance constraints being previously set meets condition；Performance in above-mentioned performance constraints
Index, it is necessary to be the numerical value class index that acquisition can be calculated from " environment ".
2. a kind of building body automatic placement method that can be constrained towards computability, which is characterized in that the moving recording, packet
It includes: the variable quantity of mobile vector, " environment " state of mobile front and back and mobile front and back performance constraints satisfaction degree.
3. a kind of groups of building automatic placement algorithm towards sunshine constraint, which is characterized in that
According to the professional knowledge in building sunshine field, " environment " number that can flexibly describe various groups of building arrangement problems is devised
It learns model (A environment), devises " agent " control module (B that can obtain the interaction experience of the single stage at this in " environment "
Agent), so that class algorithm can be learnt according to existing deeply, constructs one and be based on " agent " and " environment " mutually
" substep bout formula " the deeply learning training of dynamic experience solves module (C substep bout formula)；Wherein, " agent " constantly exists
It each obtains the state of " environment " in " single stage ", after the mobile decision function that input control is individually built, drives its movement, and
The global sunshine constraint satisfaction degree variable quantity of mobile front and back, and new " environment " state are collected, by passing through to these interactions
The movement decision function is continued to optimize in the record tested and training, is constrained to reach global sunshine in subsequent bout and step
The top of satisfaction, until completing the solution of sunshine arrangement problem.
4. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 3, which is characterized in that described to retouch
State " environment " mathematical model of groups of building arrangement problem:
By any one " the groups of building automatic placement towards sunshine constraint " problem, it is described as three based on three-dimensional system of coordinate
" environment " mathematical model is tieed up, is partially used in solution procedure for subsequent B " agent "；It includes following 5 kinds of objects:
1) sunshine measuring point: one group of three-dimensional point { P1,P2,P3..., the form to discrete three-dimensional point describes those and needs to consider day
According to the area of space of constraint；
2) sunshine measurement direction: one group of n trivector describes continuous sunshine process in one day to discrete form；Each
Vector all corresponded to one it is specific at the time of, the direction of the sun is directed toward from sunshine measuring point；
3) sunshine constraint condition parameter: one group of numerical parameter associated with each sunshine measuring point and logic judgment rule, it will be each
The different sunshine constraint condition of city, all types of buildings carries out mathematical description；
4) building: one has the geometry polyhedron of multinomial attribute data, and building for sunshine can be blocked in environment to describe
Build object entity part；
5) neighbour boundary: tier building can arrange in description neighbour and the boundary of road, neighbour range boundary, neighbour Nei Gao
The boundary of range boundary that layer building can arrange, the range that do not allow to arrange in neighbour.
5. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 3, which is characterized in that the number
Learn description, comprising:
Correspondence time interval between adjacent sunshine measurement direction: numerical value shape parameter, it describes the adjacent sunshine test of any two
The interval of time representated by direction；
The constraint satisfaction Rule of judgment of sunshine measuring point: it for one group of logic judgment rule, is built for where different cities, sunshine measuring point
Different sunshine constraint conditions of the type corresponding to it are built, describe whether sunshine constraint meets in the form of rules.
6. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 3, which is characterized in that described to build
Building object attribute includes: high-rise attribute, removable attribute, sunshine measuring point list attribute, minimum level boundary.
7. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 3, which is characterized in that described
" agent " is completed in the single stage, the control to any one building to be arranged, packet by interacting with " environment "
Contain: calculating global sunshine constraint condition satisfaction degree, select current building to be moved, calculate and observed centered on currently building
" environment " state for obtaining currently builds the movement in the case where giving vector in executions " environment ".
8. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 3, which is characterized in that described
" substep bout formula ": the interactive process between building " environment " and " agent " with u " bout ", every bout v " step "；
Period, each " bout " are restored to the original state of " environment " when starting, and in either step t, " agent " exists
In " environment " accumulation in relation to specific behavior condition, behaviour decision making, Behavioral effect interaction experience sample；When interaction experience sample number
After amount reaches requirement, in each step, so that it may be used to the deep neural network in the behavioral decision function of " agent "
It is trained；After several " bouts " several " steps ", when specific benefit, which reaches solution, to be required, solves and complete.
9. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 8, which is characterized in that set:
Specific behavior executes " environment " state the tensor St and St+1 of front and back,
By " agent " behavioral decision function generate specific behavior control tensor A t,
" agent " behavior executes the specific benefit variable quantity Rt+1-Rt of front and back " environment ".
10. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 9, which is characterized in that use
DDPG is the embodiment algorithm of deeply study,
Solution procedure includes the nested circulation for having v " step " in u " bout " and every bout, i.e., in any e bout
Any t step (e ∈ [0, u-1], t ∈ [0, v-1]) have:
Step 1: setting solves hyper parameter
According to available computational resources, various solution hyper parameters: the rounds u of calculating process, every bout number of steps v, depth are set
Neuronal quantity, connection type, activation primitive type in neural network framework；
Step 2: judge the termination condition of bout circulation
If e=u, i.e. this bout the last one bout for having exceeded setting, and before also without occurring that solution can be used, then this wheel is asked
Solution is with failure terminating；
Step 3: start a new bout k
The building in environment is subjected to initial layout by setting rule, and step counter t is reset；
Step 4: the termination condition of judgment step circulation
If t=v, i.e., this step has exceeded the last one step in the bout of setting, then bout counter e adds " 1 " (e=e+
1), and step 2 is returned；
Step 5: start the interaction step t of one new " agent " and " environment "
Step 6: global sunshine constraint satisfaction percentage Rt is calculated
By making ray along n different sunshine measurement directions respectively, whether judging it from each sunshine measuring point in environment
Geometrical intersection relationship is formed with the building in environment, i.e., analog, which calculates all sunshine measuring points, can actually obtain sunshine
Details；Further according to sunshine constraint condition parameter, can count meet sunshine constraint sunshine measuring point quantity account for owning
The ratio of sunshine measuring point quantity, i.e., global sunshine constraint satisfaction percentage Rt when step t starts；
Step 7: current building to be moved is selected
Current to be moved building of the strategically selected one removable building b as this step in " environment "；It is all
The calculating of " environment " state Sbt when the subsequent t to current procedures is all centered on it；
Step 8: global sunshine resource digraph is constructed
Construct n hiding relation digraph { G1, G2...Gn } of all sunshine measuring points on n sunshine measurement direction；Wherein,
Subgraph Gn is constructed by the following method: the floor projection vector of the sunshine measurement direction corresponding along figure traverses in scene and owns
The relationship of sunshine measuring point and neighboring buildings；It is using sunshine measuring point as the node of figure, using sunshine hiding relation as the side of figure, with from quilt
The sunshine measuring point that the sunshine measuring point blocked is possessed to the building for blocking it is positive direction；When in figure a measuring point node i with
One, its day side building forms maximum hiding relation (rise is than maximum), then all measuring points on the measuring point and the building
(such as measuring point node j) forms neighbouring relations in figure and (is such as directed toward node j) from node i；In rise ratio herein, distance
Dij is the ground line from building where the sunshine measuring point node i to adjacent node j being blocked, constituting hiding relation to it
Segment length；Height difference Hij be from it is where the sunshine measuring point node i to adjacent node j being blocked, constitute hiding relation to it and build
It builds, the vertical range at the practical top point of contact under direction sun three-dimensional ray；
Step 9: day side sunshine measuring point node is extracted
In each digraph Gn, respectively currently to move c possessed sunshine measuring point node { P1, P2...Pc } of building
For starting point, along its direction that is blocked, find it is all be located at its day side, in the building that direct or indirect hiding relation is constituted to it
Other sunshine measuring point nodes；
Step 10: it calculates day side " sunshine resources index "
For the above-mentioned node extracted from each digraph Gn, calculate wherein from sunshine measuring point corresponding to each node to
Between building corresponding to its day side adjacent node, ground distance Di on corresponding sunshine measurement direction and height difference is blocked
Hi, then be calculated as follows out current building on current digraph Gn day side " sunshine resources index (+Solar Index ,+
SI) " (formula 1)；Finally, n+SI will be obtained for n sunshine measurement direction；
Step 11: it extracts and carries on the back positive side sunshine measuring point node
In each digraph Gn, respectively currently to move c possessed sunshine measuring point node { P1, P2...Pc } of building
For starting point, inversely form the directions blocked along to other buildings, find it is all be located at its and carry on the back positive side, by building where it directly or
The sunshine measuring point node for the other buildings blocked indirectly；
Step 12: it calculates and carries on the back positive side " sunshine resources index "
For the above-mentioned node extracted from each digraph Gn, calculate wherein from sunshine measuring point corresponding to each node to
Between building corresponding to its day side adjacent node, ground distance Di on corresponding sunshine measurement direction and height difference is blocked
Hi, then be calculated as follows out current building on current digraph Gn back sun side " sunshine resources index (- Solar Index ,-
SI) " (formula 2)；Finally, n-SI will be obtained for n sunshine measurement direction；
Step 13: it calculates each to movable distance
Them are calculated separately from the centroid currently built along the floor projection of n+m both forward and reverse directions (n is sunshine measurement direction)
It arrives in the distance on corresponding neighbour boundary and (is calculated according to the corresponding neighbour boundary of the high-rise Attributions selection of building), it is practical
Transportable distance (deducting by the occupied line segment overall length of the minimum level bounds of building therebetween)；2n is obtained in total
+ 2m distance value D；
Step 14: the state Sbt of synthesis " environment "
Above-mentioned all+SI ,-SI, D, constitute the array with 4n+2m element, i.e., in the form of fixed quantity parameter
Describe the global context state Sbt observed centered on currently building；
Step 15: action vector Abt is generated with decision function
After Sbt is inputted the adjustable mobile decision function Fae of an inner parameter, it is current to calculate the vector Abt currently moved
Valuation (formula 3)；Here, as mobile decision function is during study (i.e. the training of deep neural network), inner parameter
Experience one, from random distribution, is arrived gradually convergent process, the movement that calculated vector can also be such that current monolithic builds, from nothing
The trial of purpose gradually tends to purposive exploration, and then obtains preferably global sunshine constraint satisfaction percentage；
Step 16: the movement currently built under given vector in " environment " is executed
Current building b is carried out by vector Abt present valuation in the environment practical mobile；
Step 17: the state Sbt+1 and sunshine constraint satisfaction percentage Rt+1 of " environment " after movement are calculated
According to the step identical as step 7-14, again centered on the building b after currently moving, neighboring buildings are observed, are calculated
The global state of new position describes Sbt+1 (be equivalent to building b and enter step the global state observed when t+1)；Meanwhile it calculating
This stylish all sunshine measuring point meets the percentage Rt+1 of sunshine constraint；
Step 18: storage interaction experience
By one list M for having a fixed capacity of above-mentioned data (Sbt, Abt present valuation, Sbt+1, Rt+1-Rt) deposit, (solution is opened
List is empty when the beginning, then constantly deposit；Most newly generated data replace wherein oldest data after being filled with)；
Step 19: interaction experience is trained and carrys out Optimal Decision-making function
If list M has expired, learning training is unfolded to mobile decision function；
Step 20: judge whether solution succeeds according to Rt+1
If judging that current Rt+1 has reached the requirement of Solve problems counterglow constraint, the current arrangement in environment is recorded as
Success solves, at this time " substep bout formula " solution procedure can successful termination (if necessary to multiple solutions for meeting sunshine constraint, then can be with
After record arrangement, termination algorithm, does not continue to execute subsequent 21st step)；
21st step: enter next step
Step counter t adds " 1 " (t=t+1), and returns to step 4；
11. a kind of groups of building automatic placement algorithm towards sunshine constraint as claimed in claim 9, which is characterized in that described
Step 19, interaction experience is trained and carrys out Optimal Decision-making function, set according to DDPG algorithm:
The control tensor (the mobile control vector individually built) of the specific behavior of all " agents " is by one group of inner parameter
Adjustable deep neural network (mobile decision function) Fae, Fat, Fce, Fct are calculated；Function Fae, Fat are according to the shape of " environment "
State tensor S calculates the control tensor A (formula 3, formula 4) of behavior；Function Fce, Fct are responsible for being counted according to state tensor S, tensor A
Value (formula 5, formula 6) after (i.e. mobile) execution of calculation behavior；Between the first two function, depth having the same between latter two function
Neural network structure is spent, and each coefficient carries out periodic synchronous (from Fae to Fat by algorithm；From Fce to Fct)；T step and t
There are the relationships of formula 7 between+1 step；Loss function (formula 8) using FCEloss as function Fce, with FAElossAs function Fae
Loss function (formula 9), carry out the training of deep neural network, finally the convergence state of available 4 decision functions；
Ab T present valuation=Fae(Sb t) (formula 3)
Ab T+1 long-term goal=Fat(Sb t+1) (formula 4)
QT present valuation=Fce(Sb t,Ab T present valuation) (formula 5)
QT+1 long-term goal=Fct(Sb t+1,Ab T+1 long-term goal) (formula 6)
QT long-term goal=Rt+1-Rt+QT+1 long-term goal* GAMMA (formula 7)
FCEloss=| QT long-term goal-QT present valuation| (formula 8)
FAEloss=-QT present valuation(formula 9).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810995301.0A CN109271683B (en) | 2018-08-29 | 2018-08-29 | Building group automatic arrangement algorithm for sunlight constraint |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810995301.0A CN109271683B (en) | 2018-08-29 | 2018-08-29 | Building group automatic arrangement algorithm for sunlight constraint |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109271683A true CN109271683A (en) | 2019-01-25 |
| CN109271683B CN109271683B (en) | 2020-03-20 |
Family
ID=65155035
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810995301.0A Active CN109271683B (en) | 2018-08-29 | 2018-08-29 | Building group automatic arrangement algorithm for sunlight constraint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109271683B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110097185A (en) * | 2019-03-29 | 2019-08-06 | 北京大学 | A kind of Optimized model method and application based on generation confrontation network |
| CN110874496A (en) * | 2020-01-20 | 2020-03-10 | 广东博智林机器人有限公司 | Building placement method and device based on reinforcement learning, storage medium and computer equipment |
| CN111090899A (en) * | 2019-11-07 | 2020-05-01 | 郑州大学 | Spatial layout design method for urban building |
| CN111177834A (en) * | 2019-12-27 | 2020-05-19 | 深圳集智数字科技有限公司 | Building automatic arrangement method, device, equipment and storage medium |
| CN111191312A (en) * | 2019-12-27 | 2020-05-22 | 深圳集智数字科技有限公司 | Method for obtaining block floor-arranging angle and related device |
| CN111523189A (en) * | 2020-04-23 | 2020-08-11 | 广东博智林机器人有限公司 | Building planning method and device, electronic equipment and storage medium |
| CN113987667A (en) * | 2021-12-29 | 2022-01-28 | 深圳小库科技有限公司 | Building layout grade determining method and device, electronic equipment and storage medium |
| CN114943112A (en) * | 2022-07-20 | 2022-08-26 | 深圳小库科技有限公司 | Automatic generation method, device, equipment and medium for arrangement scheme along building line |
| CN116167148A (en) * | 2023-04-26 | 2023-05-26 | 青岛理工大学 | Urban neighborhood form optimization method and system based on local microclimate |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180137214A1 (en) * | 2016-11-14 | 2018-05-17 | Autodesk, Inc. | Generative design for architecture |
| WO2018112533A1 (en) * | 2016-12-20 | 2018-06-28 | The Institute of Digital Design Australia Pty Ltd | System for determining ventilation in a specified area |
-
2018
- 2018-08-29 CN CN201810995301.0A patent/CN109271683B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180137214A1 (en) * | 2016-11-14 | 2018-05-17 | Autodesk, Inc. | Generative design for architecture |
| WO2018112533A1 (en) * | 2016-12-20 | 2018-06-28 | The Institute of Digital Design Australia Pty Ltd | System for determining ventilation in a specified area |
Non-Patent Citations (5)
| Title |
|---|
| 孙澄宇等: "面向应用的深度神经网络图说", 《时代建筑》 * |
| 宋小冬等: "日照标准约束下的建筑容积率估算方法探讨", 《城市规划汇刊》 * |
| 宋小冬等: "高层_高密度_小地块条件下建筑日照二级间距的控制与协调", 《城市规划学刊》 * |
| 董超俊等: "强化学习的基本原理", 《城市区域智能交通控制模型与算法》 * |
| 高菲: "基于日照影响的高层住宅自动布局", 《中国优秀硕士学位论文全文数据库（电子期刊）工程科技Ⅱ辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110097185A (en) * | 2019-03-29 | 2019-08-06 | 北京大学 | A kind of Optimized model method and application based on generation confrontation network |
| CN110097185B (en) * | 2019-03-29 | 2021-03-23 | 北京大学 | Optimization model method based on generation of countermeasure network and application |
| CN111090899A (en) * | 2019-11-07 | 2020-05-01 | 郑州大学 | Spatial layout design method for urban building |
| CN111177834A (en) * | 2019-12-27 | 2020-05-19 | 深圳集智数字科技有限公司 | Building automatic arrangement method, device, equipment and storage medium |
| CN111191312A (en) * | 2019-12-27 | 2020-05-22 | 深圳集智数字科技有限公司 | Method for obtaining block floor-arranging angle and related device |
| CN110874496A (en) * | 2020-01-20 | 2020-03-10 | 广东博智林机器人有限公司 | Building placement method and device based on reinforcement learning, storage medium and computer equipment |
| CN111523189A (en) * | 2020-04-23 | 2020-08-11 | 广东博智林机器人有限公司 | Building planning method and device, electronic equipment and storage medium |
| CN111523189B (en) * | 2020-04-23 | 2023-04-07 | 广东博智林机器人有限公司 | Building planning method and device, electronic equipment and storage medium |
| CN113987667A (en) * | 2021-12-29 | 2022-01-28 | 深圳小库科技有限公司 | Building layout grade determining method and device, electronic equipment and storage medium |
| CN114943112A (en) * | 2022-07-20 | 2022-08-26 | 深圳小库科技有限公司 | Automatic generation method, device, equipment and medium for arrangement scheme along building line |
| CN116167148A (en) * | 2023-04-26 | 2023-05-26 | 青岛理工大学 | Urban neighborhood form optimization method and system based on local microclimate |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109271683B (en) | 2020-03-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109271683A (en) | A kind of groups of building automatic placement algorithm towards sunshine constraint | |
| Noroozi et al. | Golden search optimization algorithm | |
| Liu et al. | Rural land use spatial allocation in the semiarid loess hilly area in China: Using a Particle Swarm Optimization model equipped with multi-objective optimization techniques | |
| CN107067091B (en) | Urban ecological land space planning model based on ant colony optimization algorithm | |
| CN108629135A (en) | Non- unified high-precision curved grid flow simulation of water quality and method for visualizing and system | |
| CN112884226A (en) | Multi-agent algorithm-based territorial spatial pattern simulation planning method and system | |
| CN109598939B (en) | A kind of prediction of short-term traffic volume method based on multitask multiple view learning model | |
| CN113051638B (en) | Building height optimal configuration method and device | |
| CN107273693A (en) | A kind of hydrocarbon fuel mechanism simplification method | |
| CN109992923A (en) | A kind of transmission line of electricity paths planning method stage by stage based on variable resolution cost surface | |
| CN104732091A (en) | Cellular automaton river bed evolution prediction method based on natural selection ant colony algorithm | |
| CN106056208A (en) | Bio-geographic optimization algorithm-oriented constraint handling method and device | |
| Sedki et al. | Swarm intelligence for groundwater management optimization | |
| Azadi et al. | GoDesign: A modular generative design framework for mass-customization and optimization in architectural design | |
| CN110222133A (en) | A kind of multiclass land use simulating and predicting method and device considering landscape index | |
| Itami | Mobile agents with spatial intelligence | |
| CN115327926A (en) | Multi-agent dynamic coverage control method and system based on deep reinforcement learning | |
| Villaggi et al. | Harnessing Game-Inspired Content Creation for Intuitive Generative Design and Optimization | |
| CN108563733B (en) | A kind of land use data assimilation method based on Bayesian inference | |
| Kumar et al. | Topological and Dimensional constraints based optimal placement of Layout Entities using Clustering and Genetic Algorithm | |
| Cheng et al. | Study of a Complete Traversal Path Planning Algorithm | |
| Rashevskiy et al. | Using Generative Design Technologies to Create Park Area Layouts for Urban Improvement | |
| CN110727291A (en) | Centralized cluster reconnaissance task planning method based on variable elimination | |
| CN117388953B (en) | Weather forecast method for improving MIM-rwkv by SADBO based on big data frame | |
| Reynolds et al. | Simulating the evolution of archaic states |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |