CN105701314A - Complex crowd evacuation behavior simulation method based on self-adaption intelligent agent model - Google Patents

Abstract

The invention discloses a complex crowd evacuation behavior simulation method based on a self-adaption intelligent agent model. The self-adaption intelligent agent model is adopted for defining four different types of attributes for each intelligent agent and can make escaping routes for each intelligent agent in a self-adaption mode according to the environment change, and therefore the self-adaption intelligent agent model has a better environmental change resisting capability. An ordinary self-adaption intelligent agent model based on a CPU platform can not deal with the complex crowd evacuation problem; a GPGPU multi-thread method based on a CUDA is adopted for developing a parallel simulation algorithm, the parallel computing thought is infused into the algorithm, and the executing efficiency of the self-adaption intelligent agent model is greatly improved while the algorithm effect is kept. Experiments prove that the method has better efficiency and usability in the actual large-scale complex crowd evacuation problem.

Description

A kind of complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model

Technical field

The invention belongs to Simulation and Modeling Technology field, relate to a kind of complicated crowd evacuation emulation method, particularly relate to a kind of complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model。

Background technology

Emergency preplan refers in the face of the accident such as contingency management of natural disaster, severe and great casualty, environmental hazard and artificial destruction, commander, rescue plan etc.。It can be effectively improved the ability of enterprise and institution's reply unexpected environmental accident, and the loss that personnel, property and environment are caused by unexpected environmental accident is minimized degree。For a long time, the formulation Dou Shi government organs of emergency preplan, social activist, the emphasis of physicist and computer scientist's research。Crowd's emergency evacuation problem is an important branch in emergency preplan, has played important effect in support personnel's safety。

The data that existing a lot of research work is all based on real world collection are analyzed ([document 1,2]) and experiment ([document 3-5]), simulation analysis is of increased attention in emergency evacuation is studied, such as, [document 6] regards a path finding problem as emergency evacuation process。Meanwhile, also having a few thing to consider the translational speed ([document 7]) of individuality, these employee's cards understand that the raising of individual translational speed can reduce the overall crowd evacuation time。In this direction, increasing work is provided to discussion decision-making mechanism in emergency evacuation and individual behavior ([document 8,9])。Additionally, also have the size that some researcheres attempt utilization measure crowd to carry out the motion track of recording individual to analyze the behavior ([document 10,11]) of crowd even with the technology of target following。Pereira et al. is proposed an evacuation management system and is predicted the mobile behavior ([document 12]) of individuality by the routing information extracted in video。

The existing evacuation model of major part and phantom are all have employed based on stream with based on individual method。Method based on stream ignores characteristic individual in crowd, thus can not consider the impact ([document 3]) that each different individual behavior is brought。Generally crowd is regarded as the overall of many individual compositions based on individual method, they be mostly based on entity or based on intelligent body (Agent)。Based in the method for entity, each individuality is modeled as not intelligent fuzzy entity, and in actual applications, can not require that all of individual behavior is all under same criterion ([document 13])。Therefore, the method based on individuality based on agent model is more favourable in actual applications, they to individuality intelligence and autonomous intelligent body (Agent) model, therefore can be individual independent carry out decision-making ([document 10])。

Model based on intelligent body (Agent) is typically concerned with the criterion defining individual behavior, and then this criterion applies to each individuality in emulation crowd。But in actual crowd evacuation process, individual behavior would generally be subject to the impact of many different factors, for instance strength, response speed, emotion and environmental effect etc. ([document 14])。These different factors affect the behavior of individuality with the form of a kind of complexity, and different individualities all can be produced different impacts by same factor。Existing certain methods can produce the simulation result that some are good, and all achieves good effect in individual and crowd behavior in actual applications。But, how individual behavior carries out adaptive adjustment remain a problem demanding prompt solution according to the change of environment and the reaction of other individualities。Therefore the emulation mode of the complicated crowd evacuation of a kind of self-adapting intelligent body Model is urgently invented。

On the other hand, the more existing method based on CPU receives great limitation in large-scale complex crowd evacuation。The existing method based on intelligent body (Agent) model can not process extensive high complexity crowd evacuation problem, (1) autgmentability of existing algorithm is very poor, along with the increase of individual number, the increase of algorithm node substantially increases the load of computing node。(2) the relatedness calculating between individuality can become more and more difficult along with the increase of individual amount。Thus urgently invent a kind of emulation mode that can quickly process large-scale complex crowd evacuation。

[document 1] P.F.Johnson, C.E.Johnson, andC.Sutherland, " Stayorgo？Humanbehavioranddecisionmakinginbushfiresandotheremergen cies, " FireTechnol., vol.48, no.1, pp.137 153,2012.

[document 2] N.ZarboutisandN.Marmaras, " Designofformativeevacuationplansusingagent-basedsimulati on, " SafetySci., vol.45, no.9, pp.920 940,2007

[document 3] R.L.Hughes, " TheFlowofHumanCrowds, " Annu.Rev.FluidMech., vol.35, pp.169 182,2003.

[document 4] D.Lee, J.H.Park, andH.Kim, " Astudyonexperimentofhumanbehaviorforevacuationsimulation, " OceanEng., vol.31, nos.8-9, pp.931 941,2004.

[document 5] Q.ZhangandB.M.Han, " Modelingandsimulationofevacuationmovementinemergency, " ProgressinSafetySci.Technol., vol.6, pp.376 380,2006.

[document 6] M.H.Zaharia, F.Leon, C.Pal, andG.PAGU, " Agent-basedsimulationofcrowdevacuationbehavior, " inProc.11thWSEASInt.Conf.Autom.Control, ModellingSimul., 2009, pp.529 533.

[document 7] H.SharbiniandA.Bade, " Modificationofcrowdbehaviormodellingundermicroscopicleve linpanicsituation, " inProc.3^rdInt.Conf.MotionGames,Nov.2010,pp.351–362.

[document 8] L.Y.JiaoandQ.Y.Jiang, " Studyonemergencymanagementinlarge-socialactivitiesbasedo nbehaviormodificationtheory, " ProgressSafetySci.Technol., vol.8, pp.451 454,2010.

[document 9] X.H.Pan, C.Han, K.Dauber, andK.Law, " Humanandsocialbehaviorincomputationalmodelingandanalysis ofegress, " Autom.Construction, vol.15, no.4, pp.448 461,2006.

[document 10] J.Berclaz, F.Fleuret, E.Turetken, andP.Fua, " Multipleobjecttrackingusingk-shortestpathsoptimization, " IEEETrans.PatternAnal.Mach.Intell., vol.33, no.9, pp.1806 1819,2011.

[document 11] F.Fleuret, J.Berclaz, R.Lengagne, andP.Fua, " Multi-camerapeopletrackingwithaprobabilisticoccupancymap, " IEEETrans.PatternAnal.Mach.Intell., vol.30, no.2, pp.267 282, Feb.2008.

[document 12] G.M.Pereira, " Dynamicdatadrivenmulti-agentsimulation, " inProc.IEEE/WIC/ACMInt.Conf.Intell.AgentTechnol., 2007, pp.76 80.

[document 13] D.Helbing, I.Farkas, andT.Vicsek, " Simulatingdynamicalfeaturesofescapepanic, " Nature, vol.407, no.6803, pp.487 490,2000.

[document 14] S.Sharma, H.Singh, andA.Prakash, " Multi-agentmodelingandsimulationofhumanbehaviorinaircraf tevacuations; " IEEETrans.Aerosp.Electron.Syst., vol.44, no.4, pp.1477 1488, Oct.2008.

Summary of the invention

For now methodical deficiency, disclosure is a kind of based on the complicated crowd evacuation parallel simulation algorithm of intelligent body (Agent)。Each individuality represents with a kind of adaptive intelligent body (Agent) model, this model adaptive in crowd evacuation process can change individual behavior, thus considering the individual character of Different Individual, this model is intended to formulate different decision-makings according to different individualities and separates the movement of different intelligent bodies。In GPU platform, algorithm is realized on an individual basis parallelization simultaneously, improves simulation algorithm crowd evacuation efficiency in large-scale complex environment。

The technical solution adopted in the present invention is: a kind of complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model, it is characterised in that comprise the following steps:

Step 1: environmental modeling；

Environmental model is a kind of multilamellar map comprising static map information and relevant macroscopic field information；Described static map information describes the physical attribute of environment；Described macroscopic field information includes a position potential energy field describing positional information and a density field describing crowd density information；Described position potential energy field is static, and density field is as the behavior of individuality and can change, and is dynamic field；

Step 2: self-adapting intelligent volume modeling；

Agent model main definitions determines the principle of intelligent body motion mode；The motion mode of intelligent body mainly includes velocity attitude and velocity magnitude；

Step 3: the parallelization of self-adapting intelligent body Model；

The parallelization of self-adapting intelligent body Model by all elements of an intelligent body queue is divided into multiple intelligent body group according to the CUDA core of GPU platform, the then each element in parallel processing intelligent body group。

As preferably, describing the physical attribute of environment described in step 1, it it is the description being carried out abstract by map；Described map is that a grid being made up of grid forms, and has the grid of four kinds of types in this grid: can pass through region grid, outlet grid, barrier grid and leader label grid；Described intelligent body by moving on region grid or outlet grid, and can only can only hold an intelligent body between each grid；After an intelligent body moves on outlet grid, it is regarded as successfully withdrawing from environment, and can be deleted from simulating scenes；All can not pass through for barrier grid intelligent body or leader label grid；The effect of described leader label grid is the positional information providing neighbouring outlet for intelligent body；

Position potential energy field described in step 1 is one and grid matrix one to one, and the corresponding specific grid of each value in matrix, this value is referred to as the position potential energy value of this grid；All of position potential energy value utilizes formula one to be standardized, and is converted into normal place potential energy value, and this value is a decimal between 0 to 1:

$p_g=\frac{p_{max}-p_c}{p_{max}-p_{min}}$ (1)

Wherein p_cIt is the current grid distance to nearest outlet grid, p_minAnd p_maxIt is the minima in the potential energy field of position and maximum respectively；P_maxValue be the position potential energy value apart from all farthest grid of any outlet；

Density field described in step 1 describes the current crowd density near ad-hoc location；Density field is one and grid matrix one to one；The corresponding specific grid of each value in matrix, this value is referred to as the density value of this grid；For specific grid k, corresponding density value ρ_kDetermined by formula two:

${\mathrm{\ρ}}_k=\frac{N_k}{N\×a^2}$ (2)

Wherein, N specifies the sum of grid in region around grid k, it is intended that region is the sub-circular region that radius is determined, N_kBeing the intelligent body sum in these grids, a is the length of side of grid。

As preferably, agent model described in step 2, intelligent body includes four class primary attributes: position attribution, Speed attribute, health degree attribute, weight properties；

Described position attribution is one group comprises the coordinate information of two values of x and y, represents grid that intelligent body currently the occupies position in map grid；

The motion mode of described Speed attribute definition intelligent body: include velocity attitude, velocity magnitude v and speed limit v_max；Described velocity attitude represents the direction that the intelligent body last time moves, and described velocity magnitude v is the intelligent body instantaneous velocity at current time, described speed limit v_maxFor the maximal rate that intelligent body can reach；

The health status of described health degree attribute description intelligent body, and when colliding, the behavior of intelligent body is had different impacts by different health status；

Each intelligent body has one group of weights, including the weights W of position potential energy_pos, density weights W_den, conformity weights W_dirWeights W with inertia_his, these weights determine behavior and the Change and Development trend of intelligent body。

As preferably, velocity magnitude v described in step 2 meets formula three:

$v=\left\{\begin{array}{c}{v}_{max},\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t)\\ v_{\max }\sqrt{\frac{{\mathrm{\ρ}}_t}{\mathrm{\ρ}}},\mathrm{\ρ}\∈\[{\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ v_{max}\sqrt{\frac{{\mathrm{\ρ}}_t{\mathrm{\ρ}}_c}{{\mathrm{\ρ}}_{max}-{\mathrm{\ρ}}_c}}\sqrt{\frac{{\mathrm{\ρ}}_{max}-\mathrm{\ρ}}{\mathrm{\ρ}}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{max})\end{array}\right.$ (3)

Wherein ρ represents the density of crowd, ρ near intelligent body_tAnd ρ_cIt is two threshold values relevant with density, ρ_maxIt is upper density limit, v_maxFor the maximal rate that intelligent body can reach；

The length of side Δ l of given actual time Δ t corresponding for each time-step and each grid, then speed v is:

$v=\frac{s}{t}=\frac{p\×\mathrm{\Δ}l}{q\×\mathrm{\Δ}t}=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}$ (4)；

Wherein, s represents that distance, t represent movement time, and time-step is the estimated time one period fixing in real world, and all of intelligent body is all move once every n time-step, and n is natural number；Integer p and q is used to calculate the intermediate value of n: n value takes the integer value closest to q/p；Therefore formula four can be expressed as formula five:

$v=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}\≈\frac{1}{n}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}$ (5)；

Determine velocity attitude by a kind of decision method based on weight described in step 2, be from 8 candidate's grids around intelligent body position, select one as next step target moved；

First calculating " the attraction force value " of eight candidate's grids around, the attraction force value of each grid is added up according to weight by the factor of position potential energy, density, conformity and four aspects of inertia；Candidate's grid is according to attracting force value sort and form a candidate queue subsequently, selects in order for intelligent body；Finally carry out cellular selection and implementation decision；

For specific intellectual Agent x and adjacent grid k, attract force value V_(k)Calculated by formula six and obtain:

V_(k)=W_pos(k)×p_g(k)+W_den×D_(k)+W_dir(k)×P_dir(k)+W_his(k)×P_his(k)(6)；

$D_{\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ \frac{{\mathrm{\ρ}}_k\left(t\right)-{\mathrm{\ρ}}_{\min }}{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_k\left(t\right)}{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.$ (7)；

$P_{dir\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.$ (8)；

$P_{his\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.$ (9)；

W_pos(k)+W_den+W_dir(k)+W_his(k)=1 (ten)；

Wherein W_pos(k)、W_den、W_dir(k)、W_his(k)It is the weight of position potential energy, density, conformity and inertia respectively；P_g(k)It it is the position potential energy value of position potential energy field；D_(k)It is that the density represented near grid k attracts force value, P_dir(k)Illustrate in the every other individuality near x, the individual shared ratio that the direction of motion is consistent with the direction of x current location to grid k；P_his(k)Represent in the motion of the nearest several times of x, the ratio that the motion consistent with the direction of motion from x to k is shared, " inertia " during namely intelligent body motion；

In formula seven, ρ_kT () represents the grid k density value in t；

In formula eight, S_dir(k)It is towards the quantity to other intelligent bodies moved with k equidirectional in the visual field of Agentx；In 8 candidate's grids, if the S that the i-th grid is corresponding_dirIt is worth maximum, then mean that major part intelligent body moves in the direction towards this grid, and Agentx is caused the trend following big group movement；

In formula nine, N_his(k)It is the nearest mobile number of times that can store of the memory of intelligent body, N_dir(k)It is that these move in number of times towards the grid k number of times moved；

Four weight sums of intelligent body are equal to 1, and these weights are the attributes that each intelligent body is inherent, thus determining the behavior of each intelligent body。

As preferably, described in carry out cellular selection and implementation decision, it implements process and is:

One intelligent body would generally move toward that grid direction attracting force value maximum, the namely best shift position of intelligent body；If best shift position can not arrive, then second-best position is selected to move；Before in enforcement, this moves, each Agentx will first calculate the feasibility of this grid, if this target side case is put shared by other Agenty, then Agentx is accomplished by being at war with Agenty；When Agentx and Agenty is at war with, the probability P that Agentx wins_swapIt is calculated according to formula 11:

$P_{swap}=H_x\×\frac{1-{\mathrm{\ρ}}_n}{1-{\mathrm{\ρ}}_x}\×\frac{{\mathrm{\ρ}}_x}{{\mathrm{\ρ}}_n}$ (11) wherein H_xIt is the health status of Agentx, ρ_xIt is the density of current grid, ρ_nBeing the density of target grid, Agentx produces the random number between a 0-1, when its size is not more than P_swapTime, Agentx and Agenty will exchange their position, say, that Agentx arrives first the position of target grid；If x has been defeated by y in competition, then x will turn to second-best grid position, by that analogy；If these 8 grids are required for competition, and agentx has lost all of competition, then Agentx just will keep current location constant。

As preferably, implementing of step 3 includes following sub-step:

Step 3.1: create a queue comprising NA intelligent body, arrange with the size ascending order of Y direction；If the size of Y direction is identical, then just carry out ascending order arrangement with the size of X-axis；Sequencer procedure carries out parallel, the size of intelligent computing agent group, k=ceil (NA/n)；Wherein n is the quantity of the stream handle of GPU platform, and k is the CUDA Thread Count in GPU；Ceil () represents a function rounded up；

Step 3.2: circulation is set and starts number of times i, i=0；

Step 3.3: retrieve i*k to (i+1) * k elementary composition i-th Agent group in intelligent body queue, and the behavior calculating each of which intelligent body makes decisions；

Step 3.4: (n-1) * k+1 to NA last elementary composition intelligent body group in retrieval intelligent body queue, then calculates the behavior of each intelligent body and makes decisions。

Traditional based on compared with the crowd evacuation emulation method of intelligent body (Agent) model with existing, the present invention has the following advantages and beneficial effect:

(1) present invention is by improving traditional agent model so that each Agent the adaptive change according to environment can be made its behaviour decision making by algorithm；

(2) each Agent is together decided on four different types of attributes the behavior of Agent so that evacuation behavior copes with the crowd evacuation problem of complexity；

(3) processed by adaptive agent model parallelization, substantially increase computing capability and the computational efficiency of model so that it is large-scale complicated crowd evacuation problem can be processed。

Accompanying drawing explanation

Accompanying drawing 1: the environmental map citing of the embodiment of the present invention。

Accompanying drawing 2: the example of the position potential energy field of the embodiment of the present invention。

Accompanying drawing 3: the example of the density field of the embodiment of the present invention。

Accompanying drawing 4: the intelligent body attribute instance figure of the embodiment of the present invention。

Accompanying drawing 5: the intelligent body behavior state transfer figure of the embodiment of the present invention。

Detailed description of the invention

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with drawings and Examples, the present invention is described in further detail, should be appreciated that enforcement example described herein is merely to illustrate and explains the present invention, be not intended to limit the present invention。

A kind of complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model provided by the invention, comprises the following steps:

Step one: environmental modeling。The present invention is described as a kind of multilamellar map containing static map information and relevant macroscopic field information environmental model。Static map information describes the physical attribute (region that such as intelligent body can pass through) of environment。Macroscopic field information includes a position potential energy field describing positional information and a density field describing crowd density information；Position potential energy field is static, and density field is as the behavior of individuality and can change, and is dynamic field；

Step 1.1: the physical attribute of environment；

The physical attribute of environment carries out the description of abstract by a map, as shown in Figure 1。Map is made up of a grid being made up of grid。This grid has the grid of four kinds of types: region (light gray), outlet (exit direction), barrier (Dark grey) and leader label (star is marked) can be passed through。Intelligent body (hereinafter referred to as Agent) by moving on region grid or outlet grid, and can only can only hold an Agent between each grid。After an Agent moves on outlet grid, it is regarded as successfully withdrawing from environment, and can be deleted from simulating scenes。All can not pass through for barrier grid Agent or leader label grid。The effect of leader label is the positional information providing neighbouring outlet for Agent。

Step 1.2: position potential energy field

Each grid in map has an associated position attribution value, and this value is determined by position potential energy field。Position potential energy field can provide the distance value of nearest outlet grid for position any given in map, and therefore Agent can rely on position potential energy field and find the optimal path of outlet recently。

Position potential energy field is a discrete 3D function, and functional image is as shown in Figure 2。Position potential energy field is usually a matrix, and the corresponding specific grid of each value in matrix, this value is referred to as the position potential energy value of this grid。The position potential energy value that barrier grid is corresponding is-2, and will be taken as cannot by treating for barrier so in the calculation。The position potential energy value of outlet grid is 0, namely minimum virtual value。Fig. 2 gives the position potential energy field of environment corresponding for Fig. 1。

In order to adapt to the decision method based on weight of Agent, all of position potential energy value to be standardized according to formula 1, is converted into normal place potential energy value, and this value is a decimal between 0 to 1:

$p_g=\frac{p_{max}-p_c}{p_{max}-p_{min}}---\left(1\right)$

Wherein p_cIt is the current grid distance to nearest outlet grid, p_minAnd p_maxIt is the minima in the potential energy field of position and maximum respectively。P_minValue be generally 0 (outlet grid to export the distance of grid be 0), p_maxValue be the position potential energy value apart from all farthest grid of any outlet。Formula result of calculation p_gBeing the normal place potential energy value that current grid is corresponding, this value is more little, and current grid is all more remote apart from any outlet；Otherwise it is then closer apart from certain outlet。Position potential energy field keeps being relatively fixed in whole simulation process, and under the premise that environment remains unchanged, potential energy value corresponding to each grid is all without changing。

Step 1.3: density field；

In map, each grid also has a corresponding density value, density field determine。Density field describes the current crowd density near ad-hoc location。Density field is the same with position potential energy field is also one and grid matrix one to one。For specific grid k, corresponding density value ρ_kDetermined by formula 2:

${\mathrm{\ρ}}_k=\frac{N_k}{N\×a^2}---\left(2\right)$

Wherein, N is the sum (specifying region is the sub-circular region that radius is determined) specifying grid in region around grid k, N_kBeing the Agent sum in these grids, a is the length of side of grid。Fig. 3 gives the example calculating density value: the density value of the five-pointed star position in figure is determined by the light gray areas that radius is 5, and this region one has N=41 grid, the length of side a=0.4 rice of each grid。A total N in this region_k=9 Agent, it is possible to calculate density value ρ_k≈ 1.372 people/square metre。

The distribution of crowd changes along with the propelling of simulation process, and therefore density field is as the motion of crowd and changing, and this point is different from static position potential energy field。

Step 2: self-adaptive agent models；

Agent model main definitions determines the principle of Agent motion mode。The motion mode of Agent mainly includes velocity attitude and velocity magnitude: wherein velocity magnitude is limited primarily by the impact of density field, and velocity attitude is then determined by a kind of decision method based on weight。

The attribute of step 2.1:Agent；

As shown in Figure 4, Agent includes four class primary attributes: position attribution, Speed attribute, health degree attribute, weight properties；Position attribution is one group comprises the coordinate information of two values of x and y, represents the Agent grid currently occupied position in map grid；Speed attribute defines the motion mode of Agent: include velocity attitude, velocity magnitude v and speed limit v_max；Described velocity attitude represents the direction that the Agent last time moves, and described velocity magnitude v is the Agent instantaneous velocity at current time, described speed limit v_maxFor the Agent maximal rate that can reach；The health degree attribute description health status of Agent, and when colliding, the behavior of Agent is had and have a great impact；Each Agent has one group of weights, including W_pos,W_dir,W_den, and W_his, these weights determine behavior and the Change and Development trend of Agent。These four weights correspondence respectively affects the four elements of Agent behavior: (1) position potential energy, (2) density, (3) conformity, (4) inertia。

The motion of step 2.2:Agent；

The motion of Agent is determined by velocity magnitude and velocity attitude。Owing to have employed the environment describing mode of similar cellular Automation Model, therefore corresponding rule is deferred in the motion of Agent: Agent moves only in the unit of whole grid, in 8 cells that every time can move adjacent to or stop motionless。

1) velocity magnitude calculates

Velocity magnitude v is determined by a piecewise function, referring to formula 3.

$v=\left\{\begin{array}{c}{v}_{max},\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t)\\ v_{\max }\sqrt{\frac{{\mathrm{\ρ}}_t}{\mathrm{\ρ}}},\mathrm{\ρ}\∈\[{\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ v_{max}\sqrt{\frac{{\mathrm{\ρ}}_t{\mathrm{\ρ}}_c}{{\mathrm{\ρ}}_{max}-{\mathrm{\ρ}}_c}}\sqrt{\frac{{\mathrm{\ρ}}_{max}-\mathrm{\ρ}}{\mathrm{\ρ}}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{max}\]\end{array}\right.---\left(3\right)$

Wherein ρ represents the density of crowd, ρ near Agent_tAnd ρ_cIt is two threshold values relevant with density, ρ_maxIt it is upper density limit。

Phantom is discrete in time and pushes ahead as unit according to time-step, one section of regular time (configurable, to give tacit consent to 0.5 second) in each time-step correspondence real world。The time system of such discrete timing be the characteristic Design according to cellular automata spatial model out, it is possible to velocity magnitude discretization thus realizing better distinguishing the purpose of speed difference between Agent in the space of cellular。In simulation process, all of Agent moves once (n is natural number) every n time-step, and the velocity magnitude v of Agent is more big, and n value is more little。For all Agent, reaching n value during maximal rate in actual emulation is 1, namely often crosses 1 time-step, Agent and moves once。Then, there is certain relation between velocity magnitude v and the n value of Agent。The length of side Δ l of given actual time Δ t corresponding for each time-step and each grid, speed v can represent with formula 4:

$v=\frac{s}{t}=\frac{p\×\mathrm{\Δ}l}{q\×\mathrm{\Δ}t}=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}---\left(4\right)$

Wherein, s represents that distance, t represent that movement time, integer p and q are used to calculate the intermediate value of n: n value takes the integer value closest to q/p, and formula 4 can be expressed as formula 5:

$v=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}\≈\frac{1}{n}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}---\left(5\right)$

2) based on set direction (decision-making) method of weight；

The important goal of Agent decision making process determines that new moving direction。In this research, Agent selects one as next step target moved based on a kind of decision method based on weights from 8 candidate's grids around oneself position。From the angle of specific Agent, first this method calculates " the attraction force value " of eight candidate's grids around, and the attraction force value of each grid is added up according to weight by the factor of aforementioned four aspect；Candidate's grid, according to attracting force value sort and form a candidate queue, is selected in order for Agent subsequently。Grid attracts force value and Agent State Transferring

For specific Agentx and adjacent grid k, attract force value V_(k)Calculated by formula 6 and obtain

V_(k)=W_pos(k)×p_g(k)+W_den×D_(k)+W_dir(k)×P_dir(k)

+W_his(k)×P_his(k)(6)

Wherein W_pos(k),W_den,W_dir(k),W_his(k)It is four factor position potential energy respectively, density, the weight of conformity and inertia。P_g(k)It is position potential energy value (formula 1), D_(k)It is that the density representing the relative density distribution near grid k attracts force value (formula 7), P_dir(k)Illustrate in the every other individuality near x, the individual shared ratio (formula 8) that the direction of motion is consistent with the direction of x current location to grid k, P_his(k)Represent in the motion of the nearest several times of x, the ratio that the motion consistent with the direction of motion from x to k is shared, " inertia " (formula 9) during namely Agent motion。

$D_{\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ \frac{{\mathrm{\ρ}}_k\left(t\right)-{\mathrm{\ρ}}_{\min }}{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_k\left(t\right)}{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.---\left(7\right)$

$P_{dir\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.---\left(8\right)$

$P_{his\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.---\left(9\right)$

In formula 8, S_dir(k)It is towards the quantity to other Agent moved with k equidirectional in the visual field of Agentx。In 8 candidate's grids, if the S that the i-th grid is corresponding_dirIt is worth maximum, then mean that major part Agent moves in the direction towards this grid, and Agentx is caused the trend following big group movement。In formula 9, N_his(k)It is the nearest mobile number of times that can store of the memory of Agent, N_dir(k)It is that these move in number of times towards the grid k number of times moved。

Four weight sums of Agent are equal to 1, as shown in Equation 10:

W_pos(k)+W_den+W_dir(k)+W_his(k)=1 (10)

These weights are the attributes that each Agent is inherent, thus determining the behavior of each Agent。

(1) if W_dir≈ 1, represents that Agent trends towards following the pedestrian of a group；

(2) if W_pos≈ 1, represents that Agent trends towards a nearest outlet；

(3) if W_den≈ 1, represents that Agent trends towards going to a less region of crowd density；

(4) if W_his≈ 1, represents that Agent trends towards keeping current direction。

The combination of four weights illustrates the state of Agent, and its state can change under given conditions, thus formulating the behavior of individuality。

In the present invention, each Agent defines four kinds of different types of behaviors, and (1) follows；(2) outlet is tended to；(3) region is tended to；(4) current direction。Fig. 5 discloses Agent motion transition diagram produced by its observation。When an Agent sees one " directional beacon ", she is likely to turn to " trend region " or " following "。When " directional beacon " disappears time, he likely turns to again " trend region " or " following " behavior。One " trend region " behavior is impossible to directly or indirectly become " following " behavior, and vice versa。

By formula (7-9), agent model can carry out the motor behavior of adaptive decision intelligent body according to the change (change of density, the change of people group, the change of surrounding) of environment。

Once the state of Agentx is determined, around it, 8 neighbour's grids attraction force value just can calculate and obtain。These 8 values are ranked up according to their size, and the moving direction of x will be the position that grid attracts force value maximum。

3) decision-making that grid selects；

After the attraction force value of 8 neighbour's grids sorts, next step is just by grid and selects and implementation decision。One Agent would generally move toward that grid direction attracting force value maximum, the namely best shift position of Agent。

If in actual applications, optimum position can not arrive people second-best position can be selected to move rather than stand in as you were。Before in enforcement, this moves, each Agentx will first calculate the feasibility of this grid。If this target side case is put shared by other Agenty, then x is accomplished by being at war with y。When Agentx and y is at war with, the probability P that x wins_swapIt is calculated according to formula 11:

$p_{swap}=H_x\×\frac{1-{\mathrm{\ρ}}_n}{1-{\mathrm{\ρ}}_x}\×\frac{{\mathrm{\ρ}}_x}{{\mathrm{\ρ}}_n}---\left(11\right)$

Wherein H_xIt is the health status of Agentx, ρ_xIt is the density of current grid, ρ_nBeing the density of target grid, Agentx produces the random number between a 0-1。When its size is not more than P_swapTime, Agentx and y will exchange their position。It is to say, x arrives first the position of target grid。If x has been defeated by y in competition, then x will turn to second-best grid position, by that analogy。If these 8 grids are required for competition, and agentx has lost all of competition, then x just will keep current location constant。

Step 3: the parallelization of self-adapting intelligent body Model。The parallelization of self-adapting intelligent body Model by all elements of an Agent queue is divided into multiple Agent group according to the CUDA core of GPU platform, the then each element in parallel processing Agent group。

Step 3.1: create the queue of a NA Agent, arrange with the size ascending order of Y direction；If the size of Y direction is identical, then just carry out ascending order arrangement with the size of X-axis。Sequencer procedure carries out parallel, calculates the size of Agent group, k=ceil (NA/n)；N is the quantity of the stream handle of GPU platform, and k is the CUDA Thread Count in GPU。

Step 3.2: circulation is set and starts number of times i, i=0；

Step 3.3: i*k to (i+1) * k elementary composition i-th Agent group in search Agent queue, and the behavior calculating each of which Agent makes decisions；

Step 3.4: (n-1) * k+1 to NA last elementary composition Agent group in search Agent queue。Then calculate each Agent behavior and make decisions。

The present invention is expanded on further below by way of experiment；

Experiment one: the effectiveness of checking self-adapting intelligent body Model。

What this experiment adopted is the emulation crowd evacuation environment shown in Fig. 1, the pedestrian of totally 1000 random distributions。Table 1 illustrates the component type of crowd, and pedestrian has attribute type four kinds different, and each type has unique H_xAnd v_max。Four key parameter W_pos,W_dir,W_den, and W_hisThe weight that each attribute is different can be adjusted by arranging its size。

Table 1AGENT type

Crowd's component type	Hx	vmax(m/s)	Quantity
				1	1.0	4	800
2	0.8	2	100
				3	0.6	1	80
4	0.4	0.5	20

By arranging different weights, carry out the Experiment Parameter of different weight as shown in table 2。Three kinds of different types of experiment (A, B, C) settings are respectively provided with as follows:

A: all except W in simulation process_pos, outside weight be both configured to 0。Experiment A is to simulate a desirable environment, is finding optimal path and is ignoring potential environmental change every time；

B: only W_dirOut in the cold, represent that the behavior of each individuality will not change along with the behavior of colony；

C: all of weight is all used, represents and adopts adaptive criterion to carry out behaviour decision making for each different individuality。

The different types of attribute weight that the different types of behavior of table 2 is corresponding

	" trend region "	" follow "	" current direction "	" tend to outlet "
					Wpos	0.7	0.5	0.91	1
Wdir	0.1	0.3	0.03	0
					Wden	0.1	0.1	0.03	0
Whis	0.1	0.1	0.03	0

Meanwhile, the ρ in step 2_tAnd ρ_cIt is arranged respectively to 0.3125 and 3.1250。Time-step in step 2 is set to 0.1s simultaneously。

The result of its three different types of experiment A, B, C is as shown in table 3 below:

3 three kinds of different types of experiments of table are in different time period remaining pedestrian's numbers

	A	B	C
				15s remains	631	655	649
30s remains	267	305	275 11 -->
				45s remains	80	134	90
60s remains	21	59	27
				Total time (s)	72.5	N/A	79.6

From table 3 it is observed that in experiment A and C, all 1000 pedestrians can evacuate complete within the time of more than 70 seconds in simulated environment。And in experiment B, have passed through the evacuation of 180s, but also still have 9 pedestrians to evacuate not yet。Can be seen that self-adapting intelligent body Model can quickly tackle complicated crowd evacuation problem。

Experiment two: the parallelization of self-adapting intelligent body Model。

What this experiment adopted is CUDA4.2 (unified calculation equipment framework), and traditional small echo coherent approach is based on VS2008C++ language and realizes。The hardware environment of experiment is NVIDIAGeForceGTX580 video card, and processor frequencies is 2004MHz, CUDA check figure order is 512。

Table 4 discloses the statistics of different number of pedestrian's number evacuation time。Common self-adapting intelligent body Model ES represents, and the result MP-ES based on GPU parallelization represents。

The statistics of the different number of pedestrian's number evacuation time of table 4

Agent number	3000	5000	8000	30000	60000
						ES(s)	1976s	4864s	10097s	NA	NA
MP-ES(s)	107s	185s	263s	971s	2330s

According to the great many of experiments in table 4, ES method evacuates 3000 people needs 1976 seconds, and the ES method after parallelization has only to 107s。ES method processes the evacuation problem of 8000 people needs 10097s, and MP-ES method has only to 263s。Present invention discover that common self-adapting intelligent body Model can only process at most the evacuation of 8000 pedestrians, and when pedestrian's quantity continues to rise, common self-adapting intelligent body Model just can not Evacuation at short notice, and the method after parallelization process is when processing the evacuation problem of 60000 people, it is only necessary to 2330s just can solve。

To sum up two experiments are it can be seen that the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model disclosed by the invention can process large-scale complicated crowd evacuation problem fast and effectively。

It should be appreciated that the part that this specification does not elaborate belongs to prior art。

Should be understood that; the above-mentioned description for preferred embodiment is comparatively detailed; therefore the restriction to scope of patent protection of the present invention can not be thought; those of ordinary skill in the art is under the enlightenment of the present invention; under the ambit protected without departing from the claims in the present invention; can also making replacement or deformation, each fall within protection scope of the present invention, the scope that is claimed of the present invention should be as the criterion with claims。

Claims (6)

1. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model, it is characterised in that comprise the following steps:

Step 1: environmental modeling；

Environmental model is a kind of multilamellar map comprising static map information and relevant macroscopic field information；Described static map information describes the physical attribute of environment；Described macroscopic field information includes a position potential energy field describing positional information and a density field describing crowd density information；Described position potential energy field is static, and density field is as the behavior of individuality and can change, and is dynamic field；

Step 2: self-adapting intelligent volume modeling；

Agent model main definitions determines the principle of intelligent body motion mode；The motion mode of intelligent body mainly includes velocity attitude and velocity magnitude；

Step 3: the parallelization of self-adapting intelligent body Model；

The parallelization of self-adapting intelligent body Model by all elements of an intelligent body queue is divided into multiple intelligent body group according to the CUDA core of GPU platform, the then each element in parallel processing intelligent body group。

2. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model according to claim 1, it is characterised in that:

Describe the physical attribute of environment described in step 1, be the description being carried out abstract by map；Described map is that a grid being made up of grid forms, and has the grid of four kinds of types in this grid: can pass through region grid, outlet grid, barrier grid and leader label grid；Described intelligent body by moving on region grid or outlet grid, and can only can only hold an intelligent body between each grid；After an intelligent body moves on outlet grid, it is regarded as successfully withdrawing from environment, and can be deleted from simulating scenes；All can not pass through for barrier grid intelligent body or leader label grid；The effect of described leader label grid is the positional information providing neighbouring outlet for intelligent body；

Position potential energy field described in step 1 is one and grid matrix one to one, and the corresponding specific grid of each value in matrix, this value is referred to as the position potential energy value of this grid；All of position potential energy value utilizes formula one to be standardized, and is converted into normal place potential energy value, and this value is a decimal between 0 to 1:

$p_g=\frac{p_{max}-p_c}{p_{max}-p_{min}}$ (1)

Wherein p_cIt is the current grid distance to nearest outlet grid, p_minAnd p_maxIt is the minima in the potential energy field of position and maximum respectively；P_maxValue be the position potential energy value apart from all farthest grid of any outlet；

Density field described in step 1 describes the current crowd density near ad-hoc location；Density field is one and grid matrix one to one；The corresponding specific grid of each value in matrix, this value is referred to as the density value of this grid；For specific grid k, corresponding density value ρ_kDetermined by formula two:

${\mathrm{\ρ}}_k=\frac{N_k}{N\×a^2}$ (2)

Wherein, N specifies the sum of grid in region around grid k, it is intended that region is the sub-circular region that radius is determined, N_kBeing the intelligent body sum in these grids, a is the length of side of grid。

3. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model according to claim 2, it is characterized in that: agent model described in step 2, intelligent body includes four class primary attributes: position attribution, Speed attribute, health degree attribute, weight properties；

Described position attribution is one group comprises the coordinate information of two values of x and y, represents grid that intelligent body currently the occupies position in map grid；

The motion mode of described Speed attribute definition intelligent body: include velocity attitude, velocity magnitude v and speed limit v_max；Described velocity attitude represents the direction that the intelligent body last time moves, and described velocity magnitude v is the intelligent body instantaneous velocity at current time, described speed limit v_maxFor the maximal rate that intelligent body can reach；

The health status of described health degree attribute description intelligent body, and when colliding, the behavior of intelligent body is had different impacts by different health status；

Each intelligent body has one group of weights, including the weights W of position potential energy_pos, density weights W_den, conformity weights W_dirWeights W with inertia_his, these weights determine behavior and the Change and Development trend of intelligent body。

4. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model according to claim 3, it is characterised in that: velocity magnitude v described in step 2 meets formula three:

$v=\left\{\begin{array}{c}{v}_{max},\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t)\\ v_{\max }\sqrt{\frac{{\mathrm{\ρ}}_t}{\mathrm{\ρ}}},\mathrm{\ρ}\∈\[{\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ v_{max}\sqrt{\frac{{\mathrm{\ρ}}_t{\mathrm{\ρ}}_c}{{\mathrm{\ρ}}_{max}-{\mathrm{\ρ}}_c}}\sqrt{\frac{{\mathrm{\ρ}}_{max}-\mathrm{\ρ}}{\mathrm{\ρ}}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{max}\]\end{array}\right.$ (3)

Wherein ρ represents the density of crowd, ρ near intelligent body_tAnd ρ_cIt is two threshold values relevant with density, ρ_maxIt is upper density limit, v_maxFor the maximal rate that intelligent body can reach；

The length of side Δ l of given actual time Δ t corresponding for each time-step and each grid, then speed v is:

$v=\frac{s}{t}=\frac{p\×\mathrm{\Δ}l}{q\×\mathrm{\Δ}t}=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}$ (4)；

Wherein, s represents that distance, t represent movement time, and time-step is the estimated time one period fixing in real world, and all of intelligent body is all move once every n time-step, and n is natural number；Integer p and q is used to calculate the intermediate value of n: n value takes the integer value closest to q/p；Therefore formula four can be expressed as formula five:

$v=\frac{p}{q}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}\≈\frac{1}{n}\×\frac{\mathrm{\Δ}l}{\mathrm{\Δ}t}$ (5)；

Determine velocity attitude by a kind of decision method based on weight described in step 2, be from 8 candidate's grids around intelligent body position, select one as next step target moved；

First calculating " the attraction force value " of eight candidate's grids around, the attraction force value of each grid is added up according to weight by the factor of position potential energy, density, conformity and four aspects of inertia；Candidate's grid is according to attracting force value sort and form a candidate queue subsequently, selects in order for intelligent body；Finally carry out grid selection and implementation decision；

For specific intellectual Agent x and adjacent grid k, attract force value V_(k)Calculated by formula six and obtain:

V_(k)=W_pos(k)×p_g(k)+W_den×D_(k)+W_dir(k)×P_dir(k)+W_his(k)×P_his(k)(6)；

$D_{\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ \frac{{\mathrm{\ρ}}_k\left(t\right)-{\mathrm{\ρ}}_{min}}{{\mathrm{\ρ}}_{max}-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{{\mathrm{\ρ}}_{\max }-{\mathrm{\ρ}}_k\left(t\right)}{{\mathrm{\ρ}}_{max}-{\mathrm{\ρ}}_{\min }},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{max}\]\end{array}\right.$ (7)；

$P_{dir\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{S_{dir\left(k\right)}}{\underset{i=1}{\overset{8}{\Σ}}{S}_{dir\left(i\right)}},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.$ (8)；

$P_{his\left(k\right)}=\left\{\begin{array}{c}0,\mathrm{\ρ}\∈\[0,{\mathrm{\ρ}}_t\]\\ 0.5\×\frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_t,{\mathrm{\ρ}}_c\]\\ \frac{N_{his\left(k\right)}}{N},\mathrm{\ρ}\∈({\mathrm{\ρ}}_c,{\mathrm{\ρ}}_{\max }\]\end{array}\right.$ (9)；

W_pos(k)+W_den+W_dir(k)+W_his(k)=1 (ten)；

Wherein W_pos(k)、W_den、W_dir(k)、W_his(k)It is the weight of position potential energy, density, conformity and inertia respectively；P_g(k)It it is the position potential energy value of position potential energy field；D_(k)It is that the density represented near grid k attracts force value, P_dir(k)Illustrate in the every other individuality near x, the individual shared ratio that the direction of motion is consistent with the direction of x current location to grid k；P_his(k)Represent in the motion of the nearest several times of x, the ratio that the motion consistent with the direction of motion from x to k is shared, " inertia " during namely intelligent body motion；

In formula seven, ρ_kT () represents the grid k density value in t；

In formula eight, S_dir(k)It is towards the quantity to other intelligent bodies moved with k equidirectional in the visual field of Agentx；In 8 candidate's grids, if the S that the i-th grid is corresponding_dirIt is worth maximum, then mean that major part intelligent body moves in the direction towards this grid, and Agentx is caused the trend following big group movement；

In formula nine, N_his(k)It is the nearest mobile number of times that can store of the memory of intelligent body, N_dir(k)It is that these move in number of times towards the grid k number of times moved；

Four weight sums of intelligent body are equal to 1, and these weights are the attributes that each intelligent body is inherent, thus determining the behavior of each intelligent body。

5. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model according to claim 4, it is characterised in that described in carry out cellular selection and implementation decision, it implements process and is:

One intelligent body would generally move toward that grid direction attracting force value maximum, the namely best shift position of intelligent body；If best shift position can not arrive, then second-best grid position is selected to move；Before in enforcement, this moves, each Agentx will first calculate the feasibility of this grid, if this target side case is put shared by other Agenty, then Agentx is accomplished by being at war with Agenty；When Agentx and Agenty is at war with, the probability P that Agentx wins_swapIt is calculated according to formula 11:

$P_{swap}=H_x\×\frac{1-{\mathrm{\ρ}}_n}{1-{\mathrm{\ρ}}_x}\×\frac{{\mathrm{\ρ}}_x}{{\mathrm{\ρ}}_n}$ (11) wherein H_xIt is the health status of Agentx, ρ_xIt is the density of current grid, ρ_nBeing the density of target grid, Agentx produces the random number between a 0-1, when its size is not more than P_swapTime, Agentx and Agenty will exchange their position, say, that Agentx arrives first the position of target grid；If x has been defeated by y in competition, then x will turn to second-best grid position, by that analogy；If these 8 grids are required for competition, and agentx has lost all of competition, then Agentx just will keep current location constant。

6. the complicated crowd evacuation behavior simulation method based on self-adapting intelligent body Model according to claim 1, it is characterised in that implementing of step 3 includes following sub-step:

Step 3.1: create a queue comprising NA intelligent body, arrange with the size ascending order of Y direction；If the size of Y direction is identical, then just carry out ascending order arrangement with the size of X-axis；Sequencer procedure carries out parallel, the size of intelligent computing agent group, k=ceil (NA/n)；Wherein n is the quantity of the stream handle of GPU platform, and k is the CUDA Thread Count in GPU；

Step 3.2: circulation is set and starts number of times i, i=0；

Step 3.3: retrieve i*k to (i+1) * k elementary composition i-th Agent group in intelligent body queue, and the behavior calculating each of which intelligent body makes decisions；

Step 3.4: (n-1) * k+1 to NA last elementary composition intelligent body group in retrieval intelligent body queue, then calculates the behavior of each intelligent body and makes decisions。