Agent-Based Systems Agent-Based Systems Where are we? Last time . . . Agent-Based Systems • Agent communication • Speech act theory Michael Rovatsos • Agent communication languages (KQML/KIF , FIPA-ACL) mrovatso@inf.ed.ac.uk • Interaction Protocols • Ontologies for communication Today . . . Lecture 7 – Methods for Coordination • Methods for Coordination 1 / 19 2 / 19 Agent-Based Systems Agent-Based Systems Methods for Coordination Coordination within interaction Coordination in a general typology of interaction: • Coordination is the process of managing inter-dependencies individual’s position between agents’ activities • Remember our previous definition isolation coexistence Coordination is a special case of interaction in which agents are aware how they depend on other agents and autosufficiency interdependence attempt to adjust their actions appropriately. • Actually this only covers agent-based coordination, but there can coordination co−action also be centralised mechanisms • In contrast to cooperation, coordination is also necessary in explicit implicit ignorance incompatibility non-cooperative systems (unless agents ignore each other) negotiation abandon goal compete 3 / 19 4 / 19
Agent-Based Systems Agent-Based Systems Typology of coordination relationships Typology of coordination relationships • Positive relationships: relationships between two agents’ plans for • More specific typology in the context of multiagent planning (von which benefit will be derived for at least one agent if plans are Martial, 1990): combined • Requests: explicitly asking for help with own activities consumable resource • Non-requested: pareto-like implicit relationships • action equality relationships: sufficient if one agent performs action resource both agents need negative non−consumable • consequence relationships: side effects of agent’s plan achieve incompatibility relationships resource other’s goals • favour relationships: side effects of agent’s plan make goal multiagent plan relationships achievement for other agent easier explicit • Basic difference to traditional computer systems: coordination is request positive achieved at run time rather than design time relationships non−requests • Remainder of lecture: discussion of different approaches to (implicit) achieve coordination 5 / 19 6 / 19 Agent-Based Systems Agent-Based Systems Partial global planning Partial global planning • Central data structure: partial global plan, containing: • Partial global planning (PGP): exchange information to reach • Objective: larger goal of the system common conclusions about problem-solving process • Activity maps: describe what agents are doing and the results of these activities • Partial – individual agents don’t generate plan for entire problem • Solution construction graph: describes how agents should interact • Global – agents use information obtained from others to achieve and exchange information to achieve larger goal non-local view of problem • Framework extended/refined in Generalized PGP (GPGP) • Three iterated stages: • GPGP introduces five techniques for coordinating activities, 1. Agents deliberate locally and generate short-term plans for goal i.e. strategies for achievement • updating non-local viewpoints (share all/no/some information) 2. They exchange information to determine where plans and goals • communicating results interact • handling simple (action) redundancy 3. Agents alter local plans to better coordinate their activities • handling hard (“negative”) coordination relationships (mainly by • Meta-level structure guides the coordination process, dictates means of rescheduling) information exchange activities • handling soft (“positive”) coordination relationships (rescheduling whenever possible, but not “mission critical”) 7 / 19 8 / 19
Agent-Based Systems Agent-Based Systems (G)PGP application – DVMT Joint intentions • We discussed intentions in practical (single-agent) reasoning • Distributed Vehicle Monitoring Testbed (DVMT): one of the earliest • But intentions also provide stability and predictability necessary for testbeds for CDPS networks social interaction • Aim of the system: tracking number of vehicles passing within a • Therefore also significant for coordination, especially teamwork range of distributed sensors • Helps to distinguish between non-cooperative and cooperative • Different problem-solving strategies were successfully tested in this coordinated activity domain using the (G)PGP approach • Basic question: in which way are individual intentions different from • Data-driven domain: challenge is to process vehicle movement (and what role do they play in) collective intentions ? data to infer their paths in a timely fashion • Remember Cohen and Levesque’s theory of intentions? They • Interesting: distributed sensor networks currently a hot topic, this extended it to teamwork situations, introducing a notion of research started in 1980! “responsibility” 9 / 19 10 / 19 Agent-Based Systems Agent-Based Systems Joint intentions Joint intentions • Example: We try to lift a stone together, and I discover it won’t work • Joint commitments have a distributed state among team members individually rational behaviour: drop the stone • Conventions describe, e.g. that an agent should inform others • However, this is not really cooperative (we should at least inform when it drops an individual commitment other) • Notion of joint persistent goal (JPG): A goal ϕ with motivation • Two important notions: (reason) ψ such that: • commitments (pledges or promises to underpin an intention) • initially all agents don’t believe ϕ but believe it is possible • conventions (mechanisms for monitoring commitment, mechanics • every agent has goal ϕ until termination condition is satisfied of adopting/abandoning commitments) • termination condition: mutual belief that ϕ satisfied, impossible to • Agents can commit themselves to actions or states of affairs achieve, or motivation ψ no longer present • While termination condition is not met, if any agent i believes ϕ is • Commitments are persistent , i.e. they are not dropped unless achieved or impossible or that ψ is no longer present it has a special circumstances arise persistent goal that this becomes mutual belief until termination • Conventions define these circumstances, e.g. that motivation for condition is met goal is no longer present, that it is or can never be achieved 11 / 19 12 / 19
Agent-Based Systems Agent-Based Systems Teamwork-based model of CDPS Mutual modelling • Based on putting ourselves in the place of the other • Practical model of how CDPS can operate using a teamwork • Involves modelling others’ beliefs, desires, and intentions . . . approach • . . . and coordinating own actions depending on resulting • Stage 1: Recognition of a goal that can be achieved through predictions cooperation (e.g. an agent can’t do it (efficiently) on his own) • Explicit communication is not necessary • Stage 2: Team formation , i.e. assistance solicitation • MACE one of the first systems to use acquaintance models for • if successful, this results in nominal commitment to collective action this purpose • deliberation phase, ends in agreement on ends (not on means) • Acquaintance knowledge involves information about others’ • rationality plays a role in deciding whether to form a group • Name unique to every agent • Stage 3: Plan formation (joint means-ends reasoning, • Class (group to which agent belongs) • Roles played by an agent in a class e.g. through negotiation or argumentation) • Skills as the capabilities of the modelled agent • Stage 4: Team action with JPG as an example convention that • Goals that the modelled agent wants to achieve governs joint plan execution • Plans describing how modelled agent attempts to achieve goals • Agent also explicitly models itself! 13 / 19 14 / 19 Agent-Based Systems Agent-Based Systems Norms and social laws Emergent social norms and laws • Example: the t-shirt game • Norms are established patterns of expected behaviour, social • agents wear red or blue t-shirt (initially at random), goal is for everyone to wear the same colour laws often add some authority to that (can be enforced or not) • agents are randomly paired in each round of the game, get to see • Idea: to strike a balance between autonomy and goals of entire other’s t-shirt colour, and then may decide to switch colour society • Problem: agent must decide which convention to adopt although • Such conventions make decision making easier for agent no global information is available • Can be designed offline or emerge from within the system • Possible update functions (=decision rules based on history): • Simple majority: agent chooses colour observed most often • The former is simpler, the latter more flexible • Simple majority with agent types: agents confide in certain other • Hard to predict which norm will be optimal for a system at design agents and exchange memory with them to inform their decision time • Simple majority with communication on success: agents will • But also hard to derive global conventions from agents’ point of communicate (successful part of) memory if success rate exceeds a threshold view given only local information • Highest cumulative reward: uses strategy that has had the highest cumulative reward so far 15 / 19 16 / 19
Recommend
More recommend
