Скачать презентацию Modeling Crowds Psycho-history Reinvented or crowd modeling and Скачать презентацию Modeling Crowds Psycho-history Reinvented or crowd modeling and

be05aa92b93ea710b59eb52380bca3c6.ppt

  • Количество слайдов: 35

Modeling Crowds: Psycho-history Reinvented (or: crowd modeling and contagion) Gal A. Kaminka The MAVERICK Modeling Crowds: Psycho-history Reinvented (or: crowd modeling and contagion) Gal A. Kaminka The MAVERICK Group Computer Science Department and Brain Research Center Bar Ilan University, Israel September 2012 Gal Kaminka [email protected] biu. ac. il

Psycho-history (in Isaac Asimov's Foundation Series) “The branch of mathematics which deals with the Psycho-history (in Isaac Asimov's Foundation Series) “The branch of mathematics which deals with the reactions of human conglomerates to fixed social and economic stimuli. ” – Gaal Dornick, “Foundation” by Isaac Asimov September 2012 Gal Kaminka [email protected] biu. ac. il

Making decisions (that affect crowds) Making decisions involves weighing uncertain outcomes To reduce uncertainty: Making decisions (that affect crowds) Making decisions involves weighing uncertain outcomes To reduce uncertainty: want predictions, and more Types of Queries: “What if” (predictions) Analyze (determine actionable factors influencing the outcomes) Plan (propose action plans to affect outcomes) • State of the art: surveys, fact-finding missions, experts, … • Limited automation Social simulation: Automation September 2012 Gal Kaminka [email protected] biu. ac. il

Social Simulation Approaches Collectives/Macro Individuals/Micro (e. g. , global dynamics) (e. g. , multi-agent Social Simulation Approaches Collectives/Macro Individuals/Micro (e. g. , global dynamics) (e. g. , multi-agent based simulation) September 2012 Gal Kaminka [email protected] biu. ac. il

Social Simulation Approaches Collectives/Macro Individuals/Micro (e. g. , global dynamics) (e. g. , multi-agent Social Simulation Approaches Collectives/Macro Individuals/Micro (e. g. , global dynamics) (e. g. , multi-agent based simulation) Qualitatively model rallies: • Predict violence • Determine actionable factors Pedestrians, evacuations: • Contagion • Culture effects [Fridman and Kaminka, SBP 2011, AMPLE 2011, QR 2011, TIST 2012] [Fridman et al. AAAI 2007, AAMAS 2012, AAMAS 2011, CSR 2011, CMOT 2010, ICCM 09, … ] September 2012 Gal Kaminka [email protected] biu. ac. il

Need Pedestrian, Evacuation Sims Training simulations – – “Urban noise” to fill virtual streets Need Pedestrian, Evacuation Sims Training simulations – – “Urban noise” to fill virtual streets Train to spot, track suspects within crowd Urban planning, architecture September 2012 Safety decision-support systems Gal Kaminka [email protected] biu. ac. il

The Goal: Individual Agent with Social Capabilities Endow each agent with capability for social The Goal: Individual Agent with Social Capabilities Endow each agent with capability for social reasoning Creates crowd phenomenon when used by many Able to account for different crowd behaviors Task independent Factors influencing action selection of agent Goal-oriented selection (most agent literature) Contagion (we do this via social comparison) Culture Emotions (e. g. , Tsai, Tambe, Marsella et al. ) September 2012 Gal Kaminka [email protected] biu. ac. il

Social Comparison Theory (SCT) (Festinger 1954) SCT: Theory in social psychology, actively researched Originally Social Comparison Theory (SCT) (Festinger 1954) SCT: Theory in social psychology, actively researched Originally given as a set of axioms (Festinger 1954) Still active research topic in psychology Key: If lacking objective means to evaluate their progress: People compare their behavior with those that are similar They take actions to reduce differences with others Tendency to reduce difference increases with similarity Hypothesis: Social comparison is the underlying mechanism of contagion 8 Gal Kaminka [email protected] biu. ac. il

SCT (Comparing agent Ame, agent set O, Similarity limits Smin, Smax) Lines 1 -4: SCT (Comparing agent Ame, agent set O, Similarity limits Smin, Smax) Lines 1 -4: Select agents not too dissimilar or too similar Line 5: Select a representative agent Ac to compare against Line 6: Determine differences with Ac Line 7: Determine power of attraction to Ac Line 8: Select an action to minimize differences 9 Gal Kaminka [email protected] biu. ac. il

Experiments: Comparison to Human Behavior Qualitative comparison: Movies of human pedestrians in Paris, Vancouver Experiments: Comparison to Human Behavior Qualitative comparison: Movies of human pedestrians in Paris, Vancouver Movies of simulated pedestrians Different variants of SCT, also non-contagion model Asked 39 subjects to rate each model: Whether it was best or worst of all models (relative fidelity) 10 How close to human (this measures absolute fidelity) Ordinal Scale: 1 (least similar) … 6 (most similar) Gal Kaminka [email protected] biu. ac. il

11 Gal Kaminka galk@cs. biu. ac. il 11 Gal Kaminka [email protected] biu. ac. il

Experiment Design Compared models from literature: Individual choice: Each agent makes decision independently SCT Experiment Design Compared models from literature: Individual choice: Each agent makes decision independently SCT with wide (2 -6. 5) and narrow (5 -6. 5), both with gain SCT with no gain, constant gains (2, 3, 4, 5) Pilot experiment threw out some of these models Subjects: 39 subjects (male 28) Movies were randomly selected From several clips of horizontal (Vancouver), vertical (Paris) From several clips of each of the simulation movies Compare horizontal to horizontal, vertical to vertical 12 Gal Kaminka [email protected] biu. ac. il

Results: Absolute Fidelity Higher results: Greater similarity to human pedestrian behavior SCT 2 -6. Results: Absolute Fidelity Higher results: Greater similarity to human pedestrian behavior SCT 2 -6. 5 significantly different than Individual and SCT 5 -6. 5 (two tailed t-test) SCT 5 -6. 5 significantly different than Individual (two tailed t-test) 13 Gal Kaminka [email protected] biu. ac. il

Results: Relative Fidelity Higher is better Lower is better 14 Gal Kaminka galk@cs. biu. Results: Relative Fidelity Higher is better Lower is better 14 Gal Kaminka [email protected] biu. ac. il

Adding Culture A Variety of documented cultural phenomena: • Passing side • Movement in Adding Culture A Variety of documented cultural phenomena: • Passing side • Movement in groups, vs. independently • Family formations • Leisurely walking speed • Personal space (proxemics) • Tendency to communicate information • Upward/downward comparison tendencies • …

A very small subset of culture results • Use webcam data, tourist videos from A very small subset of culture results • Use webcam data, tourist videos from various locations • England, France, Israel, Iraq, Canada • Measured mean parameters based on data • Were able to show good fidelity of simulation • Also, simulated mixed-culture crowds September 2012 Gal Kaminka [email protected] biu. ac. il

Modeling macro phenomena View of psycho-history: “Implicit […] is the assumption that the human Modeling macro phenomena View of psycho-history: “Implicit […] is the assumption that the human conglomerate […] is sufficiently large for valid statistical treatment. ” – Gaal Dornick, “Foundation” by Isaac Asimov September 2012 Gal Kaminka [email protected] biu. ac. il

Modeling Demonstrations, Rallies, … Goals: • Predict violence level (none, property damage, casualties) • Modeling Demonstrations, Rallies, … Goals: • Predict violence level (none, property damage, casualties) • Assist police decision making process Constraints • Expert knowledge not accurate nor complete • Mostly partial macro-level qualitative descriptions • Simulation is of large groups Proposal: Use QR (qualitative reasoning) modeling September 2012 Gal Kaminka [email protected] biu. ac. il

Qualitative Reasoning (QR) [Kuipers AIJ 84, 86, Forbus AIJ 84] • Ordinal variables: qualitative Qualitative Reasoning (QR) [Kuipers AIJ 84, 86, Forbus AIJ 84] • Ordinal variables: qualitative values rather than real numbers • Monotonic functions (increasing/decreasing, derivatives) • Algorithms simulate how variables affect each other • With partial and imprecise information • Draw useful qualitative conclusions • Physics, economics, … September 2012 Gal Kaminka [email protected] biu. ac. il

Base Model Fear of Punishment Willing Personal Price - Hostility for the police + Base Model Fear of Punishment Willing Personal Price - Hostility for the police + + Group Cohesiveness + History of Violence + Violence September 2012 Gal Kaminka [email protected] biu. ac. il

Qualitative Simulation Develops all possible behaviors from initial conditions Input: Initial state of the Qualitative Simulation Develops all possible behaviors from initial conditions Input: Initial state of the world • Contains a structural description of the model Output: State transition graph • Captures the set of all possible behaviors • developed from initial state September 2012 Gal Kaminka [email protected] biu. ac. il

What are the influencing factors on violence level? • Several theories regarding influencing factors What are the influencing factors on violence level? • Several theories regarding influencing factors • Each theory: focuses on a sub-set of factors • Challenge: combine all of them to one model • To address this challenge: • Israeli police initiated a comprehensive study, based on: • database of 102 demonstrations • interviews with 87 officers • Result: report which provides collection of factors and their influences We use this report as source of knowledge To develop QR models which enable reasoning September 2012 Gal Kaminka [email protected] biu. ac. il

Comparison of following models: • Base model: Based on literature review provided to us Comparison of following models: • Base model: Based on literature review provided to us • • By Israeli Police model • Extension of the Base model • Based on the review conclusions • Bar Ilan model • Extension of the Israeli Police model • Based on consultations with social and cognitive scientists September 2012 Gal Kaminka [email protected] biu. ac. il

Base Model Population Fear of Punishment Willing Personal Price Hostility for the police Group Base Model Population Fear of Punishment Willing Personal Price Hostility for the police Group Cohesiveness History of Violence Population Violence September 2012 Gal Kaminka [email protected] biu. ac. il

Israeli Police Model Population Group Cohesiveness Number Participants License United identity Personal Price Hostility Israeli Police Model Population Group Cohesiveness Number Participants License United identity Personal Price Hostility for the police Punishment History of Violence Group Speaker Violent core Demonstrat ion purpose Population Violence Police Time intervention Environment Intervention strength September 2012 Weather Time sensitivity Place sensitivity Gal Kaminka [email protected] biu. ac. il

Bar Ilan Model Population Personal Price License Group Speaker Number Participants Order Hostility for Bar Ilan Model Population Personal Price License Group Speaker Number Participants Order Hostility for the police History of Violence Punishment United identity Group Cohesiveness Anonymity Violent core Visual cohesiveness Demonstrat ion purpose Population Violence Police Time intervention Environment Intervention strength September 2012 Weather Time Light Place sensitivity Time sensitivity Gal Kaminka [email protected] biu. ac. il

Query 1: Predictions Compared the models on 24 real-life events • 20 demonstrations in Query 1: Predictions Compared the models on 24 real-life events • 20 demonstrations in Israel (wikipedia, in Hebrew) • 3 reported riots, with expert analysis: • Violence in Heysel stadium (1985, reported in Lewis 1989) • Los Angeles riots (1992, reported in Useem 1997) • London riot (1990, reported in Stott and Drury 2000) • Calm protest • Petach Tikva (Israel) protest (2009, video taped by us) September 2012 Gal Kaminka [email protected] biu. ac. il

Typical Qualitative Simulation Graphs Base model Police model Likelihood of each outcome: BIU model Typical Qualitative Simulation Graphs Base model Police model Likelihood of each outcome: BIU model (#behavior paths to specific outcome) (total #paths September 2012 Gal Kaminka [email protected] biu. ac. il

Subset of results: prediction accuracy The results show: 1. Police model provides poor results Subset of results: prediction accuracy The results show: 1. Police model provides poor results in prediction of Exp 4 2. Base model and BIU model provide good results in all examined cases September 2012 Gal Kaminka [email protected] biu. ac. il

Results • Level 1 errors: Off by one level • Level 2 errors: Off Results • Level 1 errors: Off by one level • Level 2 errors: Off by two levels September 2012 Gal Kaminka [email protected] biu. ac. il

Experiment 2: Sensitivity Analysis Expert analysis of reported cases: • Police used too much Experiment 2: Sensitivity Analysis Expert analysis of reported cases: • Police used too much (case 1, 2), or • too little (case 3) force. Overall, BIU model changes classification when police strength is changed September 2012 Gal Kaminka [email protected] biu. ac. il

Sensitivity Analysis (more results) Changed “police strength” variable in all 24 cases: • Police Sensitivity Analysis (more results) Changed “police strength” variable in all 24 cases: • Police model: distribution change in 3 cases, outcome change in 2 • BIU model: distribution change in 24 cases, outcome change in 7 Compared to decision-tree learning: • Use Weka J 48 (C 4. 5) for learning • Variables as attributes, so learning DTs for Police model, for BIU model • Use all 24 cases for learning (specialization is a conservative assumption) • 100% accurate on original cases • Outcome change in 3 cases (no distribution) September 2012 Gal Kaminka [email protected] biu. ac. il

Analysis query: What affects outcome? • Only subset of variables are actionable • • Analysis query: What affects outcome? • Only subset of variables are actionable • • Cannot change weather Can change police strength used • Want to know what actionable variables affect outcome • And when, under what set of conditions Algorithm analyzes simulation graph: • Find nodes with high entropy over outcomes • i. e. , nodes in which outcome is uncertain yet • Contrast variables in node and in children • Determine variable changes that shift outcome September 2012 Gal Kaminka [email protected] biu. ac. il

Analysis query: Results • Partial agreement between algorithm and experts • Algorithm does not Analysis query: Results • Partial agreement between algorithm and experts • Algorithm does not contradict the experts • • Algorithm specifies settings in which actions should be taken Experts accounted for general conditions September 2012 Gal Kaminka [email protected] biu. ac. il

Conclusions • There are different queries, that build on each other • • • Conclusions • There are different queries, that build on each other • • • Prediction: Analysis: Plan: agent-based simulation, qualitative modeling ? • Key obstacles to progress: • • • No (open) repository of data Need for interdisciplinarity No institutionalized, or funder-guided technology transfer process September 2012 Gal Kaminka [email protected] biu. ac. il