1 / 37

Multi-Agent Systemen Architecturen

Multi-Agent Systemen Architecturen. Reactive versus Hybrid agents. Reactive : no plans, thinking, memory. Only reaction; Hybrid : There are reactive and other types of control layers. Other layers may be planning, reasoning, belief, learning.

keegan
Download Presentation

Multi-Agent Systemen Architecturen

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Multi-Agent Systemen Architecturen dr. ir. M. Maris, Multi-agent systems

  2. Reactive versus Hybrid agents • Reactive: no plans, thinking, memory. Only reaction; • Hybrid: There are reactive and other types of control layers. Other layers may be planning, reasoning, belief, learning. dr. ir. M. Maris, Multi-agent systems

  3. Typical reactive agent control structure:Subsumption architecture • Use of “Behaviors” • Vertical modules • Interactions are fixed and defined by priorities of the behavioral modules • The modules can inhibit other modules Behavior x (Move to light) - Behavior 2 (Avoid obstacles) Output: Action to be executed - Behavior 1 (Grap cola can) Brooks, 1986 dr. ir. M. Maris, Multi-agent systems

  4. Other reactive control architectures • Mars explorer (Steels) • PENGI (Agre and Chapman) • Situated Automata (Rosenschein and Kaelbling) • Ant-based dr. ir. M. Maris, Multi-agent systems

  5. Limitations of reactive agents • There are no internal models, all necessary information must be extracted by the agents from their environment; • Agents can only take a short-term view (no-planning, belief); • They cannot learn from experience; • Intelligence must emerge from the interaction with the environment and that is hard to engineer and hard to understand. Neither is there any design methodology, only trial and error (examples: ants, cleaning robots); • The subsumption architecture is difficult to use if there are many (reactive) behaviors dr. ir. M. Maris, Multi-agent systems

  6. Hybrid agent-architecture Nth intelligent layer 1st intelligent layer Reactive layer dr. ir. M. Maris, Multi-agent systems

  7. Hybrid agents: Typical internal agent modules or layers • Perception / motor • Communication • Basic model of environment and other agents • Knowledge and experience, believes • Planning • Task performing • Reactive dr. ir. M. Maris, Multi-agent systems

  8. Hybrid agents: Layers of control dr. ir. M. Maris, Multi-agent systems

  9. Example hybrid architecture:Reactive with horizontal layering Agent Internal representations Decision making Planning Perception Action(s) Reactive dr. ir. M. Maris, Multi-agent systems

  10. Example2, horizontal layering:Touring Machines (by Innes Ferguson) dr. ir. M. Maris, Multi-agent systems

  11. Example two-pass control vertical layering:InteRRap (by Jörg Müller) dr. ir. M. Maris, Multi-agent systems

  12. Blackboard architecture The architecture is based on individual modules that communicate not directly but by using a “blackboard” KS1 Environment Black- board KS2 KS3 control • There are three main subsystems: • Knowledge Sources (KSs) • The shared blackboard, used by the Knowledge Sources • Control device for managing conflicts of access (compare with Wiki) dr. ir. M. Maris, Multi-agent systems

  13. Competitive tasks Task selector Task x (Move to light) Sensors Task 2 (Avoid obstacles) Actuators Task 1 (Grap cola can) Task switch dr. ir. M. Maris, Multi-agent systems

  14. Production systems If <list of conditions> then <list of actions> agent Rule base Inference engine Execution Perception Database Environment dr. ir. M. Maris, Multi-agent systems

  15. Connectionist architectures • Based on the metaphor of the brain • Using “neurons” as nodes (neural networks) • Are adaptive by changing the “weight” between two neurons • Many types of networks dr. ir. M. Maris, Multi-agent systems

  16. Connectionist architectures (2) Example of a three-layer network Out In Internal (hidden) layer output layer input layer dr. ir. M. Maris, Multi-agent systems

  17. Dynamic systems architectures The new action depends on the old action New action = old action + x Or: Action(t+1) = Action(t) + x This method tries to model the underlying agent system as accurate as possible, including real-time aspects dr. ir. M. Maris, Multi-agent systems

  18. Multi-agent Systems Utility dr. ir. M. Maris, Multi-agent systems

  19. Typical multi-agent system dr. ir. M. Maris, Multi-agent systems

  20. The idea of Utility Utility is at the individual agent level and serves the goals of that agent dr. ir. M. Maris, Multi-agent systems

  21. Preference and Utility Let’s name utility for agent i : ui Let there be a number of states w that an agent can be in. And W is the set of possible states: W = {w1 , w2 , …..} Then: ui(w) is the utility for the agent i in state w Usually, some states are preferred above other states, or: ui(w1) > ui(w2) (or, short form, w1 > w2 ) dr. ir. M. Maris, Multi-agent systems

  22. Preference and Utility • Reflexivity: for all wW it holds that w1iw2 • Transitivity: if wiw2 and w2i3 than w1i3 • Comparability: for all wW it holds that either w1iw2 or w2 iw1 So the w’s in W can be put into a certain order of preference, with properties: dr. ir. M. Maris, Multi-agent systems

  23. Utility in multi-agent systems Suppose that each agent can either do action D (defect) or C (cooperate): Ac = {C,D} Now define a transformation function t that maps the actions of the participating agents on a set of environmental states W: t: Ac (agent i) x Ac(agent j)  W dr. ir. M. Maris, Multi-agent systems

  24. Example of utility in a MAS Env 1: t(D,D) = w1; t(D,C) = w2; t(C,D) = w3; t(C,C) = w4 Env 2: t(D,D) = w1; t(D,C) = w1; t(C,D) = w1; t(C,C) = w1 Env 3: t(D,D) = w1; t(D,C) = w2; t(C,D) = w1; t(C,C) = w2 For different environment, the actions C and D may result in differences in outcome (2 agents): dr. ir. M. Maris, Multi-agent systems

  25. Numeric example of utility in MAS Suppose that: ui(w1) = 1, ui(w2) = 1, ui(w3) = 4, ui(w4) = 4 uj(w1) = 1, uj(w2) = 4, uj(w3) = 1, uj(w4) = 4 We now write: w1=t(D,D) = (D,D); w2= (D,C); w3= (C,D); w4= (C,C) and so: dr. ir. M. Maris, Multi-agent systems

  26. Numeric example of utility in MAS ui(D, D) = 1; ui(D, C) = 1; ui(C, D) = 4; ui(C, C) = 4 uj(D, D) = 1; uj(D, C) = 4; uj(C, D) = 1; uj(C, C) = 4 It follows that agent’s i preference is : C,D i C,C i D,D i D,C and agent’s j: D,C j C,C j D,D j C,D dr. ir. M. Maris, Multi-agent systems

  27. Utility in a Payoff Matrix i defects i cooperates 1 4 j defects 1 1 1 4 j cooperates 4 4 top-right is payoff received by agent i bottom-left is payoff received by agent j dr. ir. M. Maris, Multi-agent systems

  28. Strategies Dominance: Create two subsets W1 and W2 out of W W1 dominates W2 if every outcome in W1 is preferred by the agent over every outcome in W2, or: 1  1 and 2  2 it holds that w1 > w2 dr. ir. M. Maris, Multi-agent systems

  29. Actions and strategies • We define actions (set of possible actions Ac) as strategies si • We can say that s1 dominates s2 if the outcomes from s1 are preferred by the agent over the outcomes of s2 • Often one strategy will dominate the other • Between agents, a Nash equilibrium can exist dr. ir. M. Maris, Multi-agent systems

  30. Nash equilibrium Two strategies s1 and s2 are in a Nash equilibrium if • Under the assumption that agent i plays s1, agent j can do no better than s2; and, • Under the assumption that agent j plays s2 agent i can do no better than to play s1 dr. ir. M. Maris, Multi-agent systems

  31. Competitive / zero-sum interactions ui(w) + uj(w) = 0 for    Example: Chess W = {wi, w2} = {win, loose} Zero-sum interactions are always competitive dr. ir. M. Maris, Multi-agent systems

  32. Prisoner’s dilemma Two man in prison (they cannot communicate with each other) are offered: • If you confess to the crime and the other doesn’t, you are free and the other will stay in jail for 3 years; • If you both confess, you both will get 2 years; • If neither confesses, both will get 1 year dr. ir. M. Maris, Multi-agent systems

  33. Payoff Matrix Prisoner Dilemma i confess (defects) i not_confess (cooperates) j confess (defects) 2 3 0 2 j not_confess (cooperates) 0 1 3 1 Utility in number of years (boek rekent dat om, p115) What is the best strategy? Nash equilibrium! dr. ir. M. Maris, Multi-agent systems

  34. Other symmetric game: Stag Hunt Your friend and you want to appear on the last day of school with green hair, as a joke. i defects i cooperates 1 0 j defects 2 1 2 3 j cooperates 0 3 dr. ir. M. Maris, Multi-agent systems

  35. Dependence relations in MAS • Independent. They don’t depend on each other (or: their utilities do not depend on each other); • Unilateral. One agent depends on the other, but not vice versa; • Mutual. They depend on each other regarding a shared goal; • Reciprocaldependence. They depend on each other regarding a certain goal. dr. ir. M. Maris, Multi-agent systems

  36. Summarizing • Utility gives a good measure of agent’s performance; • Utility helps the agent to order the preferences of its states; • Utility supports Multi-agent systems to reach the shared goal(s) dr. ir. M. Maris, Multi-agent systems

  37. Volgende week Bayesiaanse netwerken dr. ir. M. Maris, Multi-agent systems

More Related