Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
CASE − Cognitive Agents for Social Environments Yu Zhang Trinity University | Laboratory for Distributed Intelligent Agent Systems 2 Outline • • • • • • Introduction CASE — Agent-Level Solution CASE — Society-Level Solution Experiment A Case Study Conclusion and Future Work Trinity University | Laboratory for Distributed Intelligent Agent Systems 3 Introduction • • • • • Multi-Agent Systems MAS for Social Simulation Research Goal Existing Approaches Our Approach Trinity University | Laboratory for Distributed Intelligent Agent Systems 4 Multi-Agent Systems Society Multi-Agent Agent KB KB KB KB KB Agents Societies HighFrequency Interactions Interactions are decentralized KB KB = Knowledge Base Trinity University | Laboratory for Distributed Intelligent Agent Systems 5 Simulating Social Environments 1998 1997 1996 1995 1992 First international conference on computer simulation and the social sciences. Hope is that computer simulation will achieve a disciplinary synthesis among the social sciences. Journal of Artificial Societies and Social Simulation first published. Santa Fe Institute becomes well known for developing ideas about complexity and studying them utilizing computer simulations of real-world phenomena. Series of workshops held in Italy and USA. Field becomes more theoretically and methodologically grounded. First ‘Simulating Societies’ workshop held. Trinity University | Laboratory for Distributed Intelligent Agent Systems 6 Research Goal Understanding how the decentralized interactions of agents could generate social conventions. Trinity University | Laboratory for Distributed Intelligent Agent Systems Current Approaches Agent Level Focuses on the self-interested agents. Society Level Focuses on static social structures. Trinity University | Laboratory for Distributed Intelligent Agent Systems 7 8 Our Approach Cognitive Agents for Social Environments Network Social Bounded Convention Rationality Perception Action Environment Trinity University | Laboratory for Distributed Intelligent Agent Systems 9 Related Work Sugarscape CASE COGENT SOAR Meso-Level Schelling’s Agent behavior Segregation realistic but not too computationally Model complex Top-Down Agent Complexity ACT-R CLARION Bottom-Up Trinity University | Laboratory for Distributed Intelligent Agent Systems 10 Outline • • • • • • Background and Objective CASE — Agent-Level Solution CASE — Society-Level Solution Experiment A Case Study Conclusion and Future Work Trinity University | Laboratory for Distributed Intelligent Agent Systems 11 Our Approach Cognitive Agents for Social Environments Network Social Bounded Convention Rationality Perception Action Environment Trinity University | Laboratory for Distributed Intelligent Agent Systems 12 Rationality vs. Bounded Rationality Rationality means that agents calculate a utility value for the outcome of every action. Bounded Rationality means that agents use intuition and heuristics to determine if one action is better than another. Trinity University | Laboratory for Distributed Intelligent Agent Systems 13 Daniel Kahneman Courtesy Google Image Trinity University | Laboratory for Distributed Intelligent Agent Systems 14 Two-Phase Decision Model Framing Evaluation Criteria Anchoring Selective Attention Phase I - Editing Phase II - Evaluation Accessibility State Similarity Decision Mode Two Modes of Function Intuition Deliberation Action Trinity University | Laboratory for Distributed Intelligent Agent Systems 15 Phase I - Editing Phase II - Evaluation Trinity University | Laboratory for Distributed Intelligent Agent Systems 16 Phase I - Editing • Framing: decide evaluation criteria based on one’s attitude toward potential risk and reward. • Anchoring: build selective attention on information. – Salience of information: A piece of information i is used C i , iI C I is used 1 Context of the current decision A set of all information – Anchored information: I*= {i | i > threshold} • Accessibility: determine state similarity only by I*. Accessibility relation st ~ sm if distancec, I * (st , sm ) threshold Current state A memory state Trinity University | Laboratory for Distributed Intelligent Agent Systems 17 Phase II - Evaluation • Intuition – If st ~ sm, the optimal decision policy *(st) and *(sm) should be close too. • Deliberation – Optimize *(st). Time discount factor *( st ) arg max E[ time i 0 reward( si ) | ], 0 1 i Expected value function A given policy Trinity University | Laboratory for Distributed Intelligent Agent Systems 18 Outline • • • • • • Background and Objective CASE — Agent-Level Solution CASE — Society-Level Solution Experiment A Case Study Conclusion and Future Work Trinity University | Laboratory for Distributed Intelligent Agent Systems 19 Our Approach Cognitive Agents for Social Environments Network Social Bounded Convention Rationality Perception Action Environment Trinity University | Laboratory for Distributed Intelligent Agent Systems 20 Social Convention A social law is a restriction on the set of actions available to agents. A social convention is a social law that restricts the agent’s behavior to one particular action. Trinity University | Laboratory for Distributed Intelligent Agent Systems Hard-Wired Design vs. Emergent Design Hard-wired design means that social conventions are given to agents off-line before the simulation. Emergent design is a run time solution that agents decide the most suitable conventions giving the current state of the system. Trinity University | Laboratory for Distributed Intelligent Agent Systems 21 Generating Social Conventions: Existing Rules Highest Cumulative Reward •An agent switches to a new action if the total payoff from that action is higher than the payoff obtained from the currently-chosen action. Simple Majority •An agent switches to a new action if they have observed more instance of it in other agents than the present action. •Not rely on global statistics about the system. •Rely on global statistics about the system. •Guaranteeing convergence in a 2-person 2-choice symmetric coordination game. •Convergence has not been proved. Trinity University | Laboratory for Distributed Intelligent Agent Systems 22 Generating Social Conventions: Our Rule Generalized Simple Majority Definition. Assume an agent has K neighbors and that KA neighbors are in state A. If the agent is in state B, it will change to state A with probability f (K A ) 1 1 e 2 ( 2 K A / K 1) . Theorem. When →, change to state A when more than K/2 neighbors are in state A, in a 2person 2-choice symmetric coordination game. Trinity University | Laboratory for Distributed Intelligent Agent Systems 23 24 Outline • • • • Background and Objective CASE — Agent-Level Solution CASE — Society-Level Solution Experiment – Evaluating the Agent-Level Solution – Evaluating the Society-Level Solution • A Case Study • Conclusion and Future Work Trinity University | Laboratory for Distributed Intelligent Agent Systems 25 Evaluating the Agent-Level Solution The Ultimatum Game The Bargaining Game Agent a I'll take x, you get 10x a gets x, b gets 10x Both get 0 Accept Agent b I'll take x, you get 10x a gets x, Accept b gets 10x Negotiate I'll take y, you get 10y Reject Reject or Run out of steps Both get 0 Trinity University | Laboratory for Distributed Intelligent Agent Systems 26 Phase I - Editing • Framing – 11 states: $0, $1, …, $10 • Anchoring – Use 500 iterations of Q-learning to develop anchored states • Accessibility – st ~ sm if distance * (st , sm ) 0 c, I Trinity University | Laboratory for Distributed Intelligent Agent Systems 27 Q-Learning • • • • Well studied reinforcement learning algorithm Converges to optimal decision policy Works in unknown environments Estimates long-term reward from experience expected discounted reward old value Q( st , at ) Q( st , at ) [rewardt max Q( st 1 , a) Q( st , at )] a max future value old value learning rate discount factor Trinity University | Laboratory for Distributed Intelligent Agent Systems Phase II - Evaluation • 1000 iterations of play with intuitive or deliberative decisions Trinity University | Laboratory for Distributed Intelligent Agent Systems 28 29 Results of the Ultimatum Game Human Players CASE Agents’ Results Rational Players Number Time Accepted Human Players’ Results & Rational Players’ Results Two-Phase Intuition Only Deliberation Only Split Value Trinity University | Laboratory for Distributed Intelligent Agent Systems 30 Results of the Bargaining Game Split Value 10 8 6 4 2 0 Iteration Negotiation Size Human Players’ Results Iteration Split Value Negotiation Size CASE Agents’ Results Iteration Iteration Results of the Bargaining Game by Human Players are with kind permission of Springer Science Trinity University | Laboratory for Distributed Intelligent Agent Systems 31 Outline • • • • Background and Objective CASE — Agent-Level Solution CASE — Society-Level Solution Experiment – Evaluating the Agent-Level Solution – Evaluating the Society-Level Solution • A Case Study • Conclusion and Future Work Trinity University | Laboratory for Distributed Intelligent Agent Systems 32 Evaluating the Society-Level Solution 2-person 2-choice symmetric coordination Game A B A 1 -1 B -1 1 Two Optimal Decisions: (A,A) and (B,B) Trinity University | Laboratory for Distributed Intelligent Agent Systems 33 Evaluating the Society-Level Solution • Intuitive and deliberative decisions • N agents (N2) with random initial state, A or B, with probability 50% • Agents connected by classic networks or complex networks • Evaluating two rules – Highest cumulative reward (HCR) – Generalized simple majority (GSM) • Performance measure: T90% – The time it takes that 90% of the agents use the same convention Trinity University | Laboratory for Distributed Intelligent Agent Systems 34 Classic Networks Complete Network KN Lattice Ring CN,K Random Network RN,P • Nodes fully connected to each other • Nodes fully connected to its K neighbors • Local clustering •Nodes connected with equal probability N=8 N=100 K=6 N=100 P=5% Trinity University | Laboratory for Distributed Intelligent Agent Systems 35 Complex Networks Small-World Network WN,K,P •Start with a CN,K graph and rewire every link at random with P •Local clustering & randomness N=100 K=6 P=5% Scale-Free Network SN,K, •P(K) is a power law P(K) ~ K- •Large networks can self-organize into a scale free state, independent of the agents N=100 K=6 =2.5 Trinity University | Laboratory for Distributed Intelligent Agent Systems 36 Evaluating Highest Cumulative Reward Network Topology SN CN KN Name Scale-free network Lattice ring Complete network Size 103 (N) 1, 2.5, 5, 7.5, 10, 25, 50 0.1, 0.25, 0.5, 0.75, 1 1, 2.5, 5, 7.5, 10, 25, 50 Parameter =2.5 <K>=12 K=12 None needed WN Small-world network 1, 2.5, 5, 7.5, 10 P=0.1 <K>=12 Lattice ring T90%=O(N2.5) Small-world T90%=O(N1.5) Scale-free/ Complete T90%=O(NlogN) Trinity University | Laboratory for Distributed Intelligent Agent Systems 37 Evaluating Generalized Simple Majority Network Topology CN KN SN Name Lattice ring Complete network Scale-free network Size 103 (N) 0.1, 0.25, 0.5, 0.75 1, 2.5, 5, 7.5, 10, 25, 50, 75, 100 1, 2.5, 5, 7.5, 10, 25, 50, 75, 100 Parameter K=12 None needed =2.5 <K>=12 SN Scale-free network 1, 2.5, 5, 7.5, 10, 25, 50, 75, 100 =3.0 <K>=12 WN Small-world network 1, 2.5, 5, 7.5, 10, 25, 50 P=0.1 <K>=12 Lattice ring T90%=O(N2.5) Small-world T90%=O(N1.5) Scale-free/ Complete T90%=O(N) Trinity University | Laboratory for Distributed Intelligent Agent Systems 38 Evaluating HCR vs. GSM Network Topology Small-World Network Lattice ring N 104 <K> 12 P P=0.05, 0.09 … 0.9 (P=0.09) Random network Trinity University | Laboratory for Distributed Intelligent Agent Systems