Presentation is loading. Please wait.

Presentation is loading. Please wait.

Zhai Yuqing Learning in Multi-agent System.

Similar presentations

Presentation on theme: "Zhai Yuqing Learning in Multi-agent System."— Presentation transcript:

1 Zhai Yuqing Learning in Multi-agent System

2 Outline Agent Learning Multi-agent learning Reinforcement Learning & Multi-agent Reinforcement Learning

3 Agent Learning

4 Why Learning? Learning is essential for unknown environments, –i.e., when designer lacks omniscience Learning is useful as a system construction method, –i.e., expose the agent to reality rather than trying to write it down Learning modifies the agent's decision mechanisms to improve performance

5 Why Learning? It is difficult to hand-design behaviours that act optimally (or even close to it) Agents can optimize themselves using reinforcement learning –Not learning new concepts (or behaviours), but given a set of states and actions it can find best policy –Is this just adaptive control? Learning can be done on-line and continuously throughout lifetime of agent, adapting to (slowly) changing situations

6 Learning Policy World, State Learning Algorithm Actions Observations, Sensations Rewards

7 Learning to act in the world Policy Environ- ment Learning Algorithm Actions Observations, Sensations Rewards Other agents (possibly learning) ? World

8 Learning Agent Architecture A learning agent can be thought of as containing a performance element that decides what actions to take and a learning element that modifies the performance element so that it makes better decisions

9 Learning Agent Architecture



12 Multi-agent Learning

13 Learning in Multiagent Systems Intersection of DAI and ML Why bring them together? –There is a strong need to equip Multiagent systems with learning abilities –The extended view of ML as Multiagent learning is qualitatively different from traditional ML and can lead to novel ML techniques and algorithms

14 Multi-Agent Learning Problem: Agent tries to solve its learning problem, while other agents in the environment also are trying to solve their own learning problems.  challenging non-stationarity. Main scenarios: (1) cooperative; (2) self-interest (many deep issues swept under the rug) Agent may know very little about other agents: –payoffs may be unknown –learning algorithms unknown Traditional method of solution: game theory (uses several questionable assumptions)

15 Multi-agent Learning Problem Agent tries to solve its own learning problem, while other agents in the environment try to solve their own learning problems Larger state space –Might have to include state of other robots in own state Problems of Multi-Agent RL –All of problems from single agent –Other agents unpredictable or non-stationary –Should reinforcement be local or global? –Was robot trying to achieve goal or reacting to other robots when doing a good action?

16 Learning in Multi-Agent Systems No Doubt learning is of great importance for MAS ! Challenge: –Multi-Agent learning problem. The optimal policy changes. Other agents are learning too. –Can we have a unifying framework in which this learning can be understood. Challenging MAS-domains: –Robotic soccer –Traffic –Robotic rescue –Trading agents, e-commerce –Automated Driving





21 General Characterization Principal categories of learning The features in which learning approaches may differ The fundamental learning problem known as the credit-assignment problem

22 Principal Categories Centralized Learning (isolated learning) –Learning executed by a single agent, no interaction with other agents –Several centralized learners may try to obtain different or identical goals at the same time

23 Principal Categories Decentralized Learning (interactive learning) same –Several agents are engaged in the same learning process –Several groups of agents may try to obtain different or identical learning goals at the same time Single agent may be involved in several centralized/decentralized learning processes at the same time

24 Learning and Activity Coordination Previous research on coordination focused on off- line design of behavioral rules, negotiation protocols, etc… Agents operating in open, dynamic environments must be able to adapt to changing demands and opportunities How can agents learn to appropriately coordinate their activities?

25 Learning about and from Other Agents Agents learn to improve their individual performance Better capitalize on available opportunities by prediction the behavior of other agents (preferences, strategies, intentions, etc…)

26 Learning Organizational Roles Assume agents have the capability of playing one of several roles in a situation Agents need to learn role assignments to effectively complement each other

27 Learning Organizational Roles The framework includes Utility, Probability and Cost (UPC) estimates of a role adopted at a particular situation –Utility –Utility – desired final state’s worth if the agent adopted the given role in the current situation –Probability –Probability – likelihood of reaching a successful final state (given role/situation) –Cost –Cost – associated computational cost incurred –Potential –Potential – usefulness of a role in discovering pertinent global information

28 Learning Organizational Roles: Theoretical Framework sets of situations and roles for agent k An agent maintains vectors of UPC During the learning phase: rates a role by combining the component measures

29 Learning Organizational Roles: Theoretical Framework After the learning phase is over, the role to be played in situation s is: UPC values are learned using reinforcement learning UPC estimates after n updates:

30 Learning Organizational Roles: Updating the Utility S – the situations encountered between the time of adopting role r in situation s and reaching a final state F with utility The utility values for all roles chosen in each of the situation in S are updated:

31 Learning Organizational Roles: Updating the Probability - returns 1 if the given state is successful The update rule for probability:

32 Learning Organizational Roles: Updating the Potential - returns 1 if in the path to the final state, conflicts are detected and resolved by information exchange The update rule for potential:

33 Learning to Exploit an Opponent: Model-Based Approach minimaxThe most prominent approach in AI for developing playing strategies is the minimax algorithm –Assumes that the opponent will choose the worst move An accurate model of the opponent can be used to develop better strategies

34 Learning to Exploit an Opponent Model-Based Approach The main problem of RL is its slow convergence Model based approach tries to reduce the number of interaction examples needed for learning Perform deeper analysis of past interaction experience

35 Model Based Approach The learning process is split into two separate stages: –Infer a model of the other agent based on past experience –Utilize the learned model for designing effective interaction strategy for the future

36 Reducing Communication by Learning Learning is a method for reducing the load of communication among agents Consider the contract-net approach: –Broadcasting of task announcement is assumed –Scalability problems when the number of managers/tasks increases

37 Reducing Communication in Contract-Net addressee learningA flexible learning-based mechanism called addressee learning Enable agents to acquire knowledge about the other agents’ task solving abilities Tasks may be assigned more directly

38 Reducing Communication in Contract-Net Case-based reasoning is used for knowledge acquisition and refinement Humans often solve problems using solutions that worked well for similar problems Construct cases – problem-solution pairs

39 Case-Based Reasoning in Contract Net Each agent maintains it own case base A case consists of: –Task specification –Info about which agent already solved the task and the quality of the solution Need a similarity measure for tasks

40 Case-Based Reasoning in Contract Net Distance between two attributes is domain-specific Similarity between two tasks and : For task a set of similar tasks is:

41 Case-Based Reasoning in Contract Net An agent has to assign task to another agent Select the most appropriate agents by computing their suitability:

42 Improving Learning by Communication Two forms of improving learning by communication are distinguished: –Learning based on low-level communication (e.g. exchanging missing information) –Learning based on high-level communication (e.g. mutual explanation)

43 Improving Learning by Communication Example: Predator-Prey domain –Predators are Q-learners –Each predator has a limited visual perception –Exchange sensor data – low-level communication –Experiments show that it clearly leads to improved learning results

44 Knowledge exchange in MAS More sophisticated implementations provide knowledge exchange capabilities – Exchange the strongest rules they have learned –Multi-agent Mutual Learning(MAML)

45 Some Open Questions… What are the unique requirements and conditions for Multiagent learning? Do centralized and decentralized learning qualitatively differ from each other? Development of theoretical foundations of decentralized learning Applications of Multiagent learning in complex real-world environments

46 Reinforcement Learning & Multi-agent Reinforcement Learning

47 6

48 12

49 Reinforcement Learning Approach Overview: TD [Sutton 88], Q-learning [Watkins 92 ] Agent can estimate a model of state transition probabilities of E (Environment), if E has a fixed state transition probability (; E is a MDPs). Profit sharing [Grefensttette 88] Agent can estimate a model of state transition probabilities of E, even though E does not have a fixed state transition probability. c.f. Dynamic programming Agent needs to have a perfect model of state transition probabilities of E. Feature:  The Reward won’t be given immediately after agent’s action.  Usually, it will be given only after achieving the goal.  This delayed reward is the only clue to agent’s learning. State Recognizer Action Selector LookUp Table W ( S, a ) Learner Agent Input Action Environment Reward Environment E








57 Agent Environment action a t r t+1 s t+1 reward r t state s t Reinforcement Learning Scenario







64 Example Q(s, a red ) = 0 +  × 81 = 72 Q(s, a green ) = 0 +  × 100 = 90 Q(s, a blue ) = 0 +  × 100 = 90












76 Multi-agent RL Basic idea –Combine the learning process in an unknown environment with the interactive decision process of multiple agents – There is no single utility function to optimize –Each agent has a different objective and its payoff is determined by the joint action of multiple agents

77 Challenges in Multi-agent RL Curse of dimensionality –The number of parameters to be learned increases dramatically with the number of agents Partial observability –states and actions of the other agents which are required for an agent to make decision are not fully observable –inter-agent communication is usually costly –Notes: Partially observable Markov decision processes (POMDPs) have been used to model partial observability in probabilistic AI

78 The Multi-Agent Reinforcement Learning (MARL) Model Multiple selfish agents in a stationary dynamic environment. Environment modeled as a Stochastic (a.k.a Markov) Game (SG or MG). Transition and Payoff are functions of all agents’ actions.

79 The MARL Model (cont.) Transition probabilities and payoffs are initially unknown to agent. Agent’s goal – maximize return.

80 Typical Multi-agent RL methods Value Iteration Learning [Sutton and Barto ] – Based on different concepts of equilibrium in game theory Min-max solution-based learning algorithm in zero- sum stochastic games [Littman] Nash Equilibrium-based learning algorithm [ Wellman ] –Extending Littman ’ s algorithm to the general-sum games Correlated Equilibrium based learning algorithm[Hall ] –Consider the possibility of action correlation among agents

81 Typical Multi-agent RL methods Multiple-person decision theory-based – Assume that each agent plays a best-response against stationary opponents – Require the joint action of agents to converge to Nash Equilibrium in self-play –Learn quickly while losing and slowly while winning –Learn best response when opponents are stationary, otherwise move to equilibrium

82 Typical Multi-agent RL methods Integrating RL with Coordination Learning –Joint action learner Independent Learner –Ignore the existence of other agents –Just apply RL in the classic sense

83 Typical Multi-agent RL methods Hierarchical Multi-agent RL –Each agent is given an initial hierarchical decomposition of the overall task –Define cooperative subtasks to be those subtasks in which coordination among agents has significant effect on the performance of the overall task –Cooperative subtasks are usually defined at highest level(s) of the hierarchy

84 MAL Foundation The game theoretic concepts of stochastic games and Nash equilibria Learning algorithms use stochastic games as a natural extension of Markov decision processes (MDPs) to multiple agents –Equilibria learners Nash-Q, Minimax-Q Friend-or-Foe-Q gradient ascent learner –best-response learner

85 Multiagent Q-learning desiderata  “ performs well ” vs. arbitrarily adapting other agents  best-response probably impossible  Doesn ’ t need correct model of other agents ’ learning algorithms  But modeling is fair game  Doesn ’ t need to know other agents ’ payoffs  Estimate other agents ’ strategies from observation  do not assume game-theoretic play  No assumption of stationary outcome: population may never reach eqm, agents may never stop adapting  Self-play: convergence to repeated Nash would be nice but not necessary. (unreasonable to seek convergence to a one- shot Nash)

86 Finding Nash equilibrium Game theoretic approach which supposes the complete knowledge of the reward structure of the underlying game by all the agents – Each agent calculates an equilibrium, by using mathematical programming – Suppose that the other agents are rational

87 Potential applications of MARL E-commerce – agents buying and selling over the internet. Autonomous computing, e.g., automatic fault recovery. Exploration of environments that are inaccessible to humans: bottom of oceans, space, etc…

88 The End

Download ppt "Zhai Yuqing Learning in Multi-agent System."

Similar presentations

Ads by Google