Multi-agent Systems, Communication and Coordination Multi-Agent Systems Course Koen Hindriks Delft University of Technology, The Netherlands Koen Hindriks, Birna van Riemsdijk Multi-agent systemen GOAL So Far • • • • • • mental states (beliefs, goals) mental state conditions & macros action specifications, built-in actions action rules environment & perception modules Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Outline (today) • Intro Multi-Agent Systems (MAS) • Coordination in MAS • Specifying a MAS in GOAL • Communication in GOAL Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Multi-Agent Systems (MAS) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Multi-Agent System Intersecting spheres of influence of multiple agents acting in an environment A1 Agents A3 A2 A4 Environment Intersecting spheres: •Action perspective: may interfere, change same parts of environment •Perception perspective: agents may have different views on environment Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Multi-Agent System Various approaches to coordinate activity of multiple agents A1 Agents A3 A2 A4 Environment Various coordination techniques have been proposed, main ideas: • Communication protocols impose structure on agent communication • Add organization structure on top of multi-agent system Note: Without explicit agent communication still implicit communication possible via environment if perception spheres overlap Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Multi-Agent System Not all agents need to be ‘connected’ to the environment A3 A1 Agents A2 A4 Environment Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Kinds of multi-agent systems • Cooperative versus competitive – Agents have a common goal of solving a problem together or compete for resources. • Homogeneous versus heterogeneous – All agents have the same abilities and decision logic, or agents have different abilities. • Interaction protocols – Which agents can talk to which other agents? • Organizational structure – Agents may be organized in various ways (with a leader, or not, different roles, etc.) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Example – Search and Rescue • Search & Rescue – Unknown environment (map may be unreliable or absent) – Search for victims – Providing aid to victims – Aim to minimize time to provide aid to all victims. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Example – Package Delivery • Package delivery – Known environment (map) – Initial and target location of packages known – Aim to minimize delivery time of all packages. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Advantages of MAS • Complete tasks quicker through shared effort. • Achieve tasks otherwise not possible. A MAS provides solutions in situations where action is spatially and temporally distributed. • A MAS can be distributed over machines across a network. • A decentralized MAS does not suffer from the "single point of failure" problem. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Cooperation & Coordination Maaike Harbers, Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Examples of Cooperation http://cacm.acm.org/news/98755-how-football-playing-robots-have-the-future-of-ai-at-their-feet/fulltext Maaike Harbers, Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Cooperation Challenges • Cooperation yields more effective (e.g., quicker) solutions to a problem… • But… 1. Avoid duplication of effort. 2. Avoid interference, i.e. harmful interactions. 3. Avoid communication overhead, i.e., need to share information with as few messages as possible. 4. Need to synchronize behaviors. • Coordination needed to address challenges Koen Hindriks, Birna van Riemsdijk Multi-agent systemen What is coordination? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Coordination problems • Without coordination, what would happen here? • What is the problem: duplication of effort, interference, communication overhead, synchronization? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Coordination problems • And what would happen here? What is the problem: duplication of effort, interference, communication overhead, synchronization? http://www.wired.com/gadgetlab/2010/06/dance-dance-revolution-20-robots-think-they-can-dance/ Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Coordination problems • And what would happen here? What is the problem: duplication of effort, interference, communication overhead, synchronization? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Coordination problems • And what would happen here? Player A / Player B Defect Cooperate Defect 2,2 5,0 Cooperate 0,5 3,3 Prisoner's dilemma Two persons are being questioned in separate rooms by the police. Each of them can either admit to a crime (defect) or not (cooperate). Outcomes for actors A and B depend on the combination of A's and B's action. What is the problem: duplication of effort, interference, communication overhead, synchronization? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Need to be smart • And what would happen here? The wise men problem A king wishing to know which of his three wise men is the wisest, paints a white spot on each of their foreheads, tells them at least one spot is white, and asks each to determine the color of his spot. … What is the problem: duplication of effort, interference, communication overhead, synchronization? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Types of coordination • Implicit coordination e.g., flocking behavior • Explicit coordination exchange of messages Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Aspects of joint activity Coordinating behavior makes multiple agents act like a team, they perform a joint activity. Klein, G., Feltovich, P., Bradshaw, J. M., & Woods, D. D. (2005). Common ground and coordination in joint activity. Organizational Simulation. W. B. Rouse and K. R. Boff. New York City, NY, John Wiley. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen A. Criteria: example Making an assignment together • No intention – no assignment or two assignment • No interdepencence – each making one half of the assignment without seeing the other half Koen Hindriks, Birna van Riemsdijk Multi-agent systemen B. Requirements for joint activity • Interpredictability – Need to be able to accurately predict what others will do – Not just actions, but the coordination itself • Common ground – Pertinent mutual knowledge, beliefs, and assumptions about others’ skills and capabilities – Joint activity is guided by signaling and coordination devices • Directability – Capacity for modifying the actions of the other parties as conditions and priorities change – Responsiveness of each participant to the influence of the others Koen Hindriks, Birna van Riemsdijk Multi-agent systemen B. Requirements: example Football team • Interpredictability – predict where team members will be on the field and who will be available • Common ground – who is keeper, who defends, who attacks • Directability – keeper tells others where to stand during a corner Koen Hindriks, Birna van Riemsdijk Multi-agent systemen C. Choreography of joint activity • Three phases (functional breakdown) • coordinating tasks is coordinating subtasks • subtasks have an entry, body and exit phase • Signaling transitions • within and between phases • Coordination devices • serve to increase interpredictability • examples are organizational structure, convention, communication protocol Koen Hindriks, Birna van Riemsdijk Multi-agent systemen C. Choreography: example Traffic • Three phase actions - entering a road being on the road leaving a road • Signaling transitions - turn signal • Coordination devices - organizational structure, convention, communication protocol to discuss… Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Coordination structures Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Organizational structure from Bradshaw et al. Koen Hindriks, Birna vanadapted Riemsdijk Multi-agent systemen Organizational structure in BW4T • Notion of ‘role’ is central • Possible roles in BW4T? • searcher • deliverer • manager Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Notion of ‘role’ is central organization goals determine agent roles performed by agents Organizational perspective Koen Hindriks, Birna van Riemsdijk agent goals determine agent roles performed within society Agent perspective Multi-agent systemen Conventions Coordination by convention solution Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Conventions in BW4T • A group of agents • Each agent is able to solve the BW4T task individually • Agents do not take into account that there are other agents • What happens? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Example of conventions in BW4T • Do not enter rooms that are occupied • If you know about a block that can be delivered, deliver that block • If someone intends to deliver a particular block, do not try to deliver that block as well Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Conventions in organizational structure If a searcher finds a block, it must report that to the manager adapted from Bradshaw et al. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Communication in BW4T • A group of agents • Each agent is able to solve the BW4T task individually • Agents take into account that there are other agents • Agents do not communicate • What happens? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Implicit coordination in BW4T • And what would happen here? Room 1 Room 2 Robot 2 Robot 1 Robots do NOT communicate. How do they choose a room? Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Specifying a MAS Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification In GOAL, a .mas file is a recipe for launching a multi-agent system, specifying: • which environment will be launched, • how many agents will be launched, • the agent names for agent processes • which agents are connected to environment • which .goal file(s) is (are) used to create an agent. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification A .mas file is a recipe for launching a MAS: • which environment will be launched environment{ env = "blocksworld.jar". init = [configuration="bwconfigEx1.txt"]. } • environment started by jar file, • some of which may be initialized using the init command (check out environment doc). Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification A .mas file is a recipe for launching a MAS: • which agent files will be used agentfiles { "coffeemaker.goal". "coffeegrinder.goal". } • agent file defines a possible agent type. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification A .mas file is a recipe for launching a MAS: • which agent processes will be created launchpolicy { launch maker:coffeemaker. launch grinder:coffeegrinder. } • a launch policy specifies when, which and how many agent processes will be created. • launch maker:coffeemaker. launches one agent process called maker using the agent file referenced by coffeemaker (coffeemaker.goal). • does not connect agent to an environment. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification A .mas file is a recipe for launching a MAS: • which agent processes will be created launchpolicy { when entity@env do launch elevator : file1. } • a conditional launch rule is triggered by the environment: whenever a controllable entity becomes available in the environment, an event is generated. • the agent created by the launch part of the rule is connected to that entity. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification A .mas file is a recipe for launching a MAS: • additional checks in launch rule launchpolicy { when [type=car,max=3]@env do launch elevator:file1. } • this launch rule can only be applied if there is an available entity of type car, and the rule can be applied at most 3 times. Remark: launch rules are applied in linear order. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen MAS Specification: Example A .mas file is a recipe for launching a MAS: agentfiles { "coffeemaker.goal". "coffeegrinder.goal". } launchpolicy { launch maker:coffeemaker. launch grinder:coffeegrinder. } • environment section may be empty, in that case mas is run without an environment. • agentfiles and launchpolicy must be non-empty. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Agent Identity and Other Agents How does an agent know it is not alone? • The names of all agents generated by GOAL are inserted in the belief base of all other agents. • For example, agent(tarzan) is inserted in the belief base of agent jane when tarzan is created. How is an agent able to differentiate itself from other agents? I.e. how does agent jane know she is not tarzan? • Upon creation, the fact me(<name>) where name is the agent’s name is inserted in the belief base of that agent. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Communication in GOAL Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Communication: Primitives • Communication primitives: – send(<recipients>, <content>) – sendonce(<recipients>, <content>) • Example: Tarzan sends Jane a message: – send(jane, meeting(date(1,2,2010), time(12,0), duration(1,0), [jane, tarzan])) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen GOAL Agent Architecture • The basis of communication in GOAL is a simple mailbox semantics. Agent message percept Process message Beliefs Process percepts Goals Action Rules Environment Real or simulated world of events Koen Hindriks, Birna van Riemsdijk Multi-agent systemen action Communication: Semantics Action: send(<recipients>, <content>) • ground send action can always be performed, i.e. precondition is true, • assume agent sender sends message to recipients • then GOAL puts sent(<recipients>, <content>) in the mailbox of sender, • agents that receive message put received(sender, <content>) NB: keep in mind it may take some time in their mailbox. before agent b receives the message. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Communication: Semantics Action: sendonce(<recipients>, <content>) • same as send(<recipients>, <content>), but • sendonce action can only be performed if the mailbox of the sending agent does not contain: sent(<recipients>, <content>) i.e. precondition is: not(sent(<recipients>, <content>) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Using Messages in Mailbox The facts in the mailbox can be used like any other facts in the belief base of an agent. That is, • sent(<recipient>, <content>) and • received(name, <content>) can be used in: – in conditions in action rules – clauses in the knowledge or beliefs section Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Using Messages in Mailbox • After receiving a message from tarzan, agent jane will have: received(tarzan, meeting(date(1,2,2010), time(12,0), duration(1,0), [jane,tarzan])) in her mailbox. • Consequently, jane will believe: bel(received(tarzan, meeting(date(1,2,2010), …, …, [jane,tarzan]))) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Processing Message Pattern • Inserting received information into belief base and removing received message: if bel(received(A,Proposition)) then insert(Proposition) + delete(received(A,Proposition)) • Removing received message cleans up the mailbox and makes sure that the action rule is not applied repeatedly (even after the received info is outdated). • Using received messages in mental state conditions other than for processing the message, and using them in knowledge base, gives less readable code. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Message Types • GOAL supports three message types, called moods: – declarative, typically used to inform Example: The time is 2 o'clock operator: ‘:’ : send(jane, :time(14,0)) – imperative, typically used to indicate a goal Example: I want the door to be closed! operator: ‘!’ : send(jane, !door(closed)) – interrogative, typically used to ask a question Example: How much milk is left? operator: ‘? : send(jane, ?amountMilk(_)) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Message Types Mood operators are translated to Prolog predicates in mailbox: – declarative: no predicate received(tarzan, time(14,0)) – imperative: imp received(tarzan, imp(door(closed))) – interrogative: int received(tarzan, int(amountMilk(_))) See programming guide, section 6.5. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Agent Selectors First argument of a send action is called an agent selector, which can be: •An agent name: send(jane, …) •A variable (to be instantiated): send(Person, …) •A list of selectors: send([jane,tarzan,Person], …) •A quantor: – send(self, …) : message sent to agent itself – send(all, …) : message sent to all agents (incl. itself) – send(allother, …) : message sent to all other agents Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Sending messages to oneself • Messages sent to oneself change the agent’s own mental state! Sending messages to oneself has special meaning. • Changes are different for different moods: – send(self, :fact) inserts fact in belief base – send(self, !fact) adopts fact as goal and deletes fact from belief base. – send(self, ?fact) deletes fact from belief base N.B.: Do not use messages to self as an alternative way of updating the agent’s beliefs and goals! Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Selecting agents to send to • Suppose agent wants to send to all other agents that are producers of milk that it wants to have milk. Q: Write an action rule which achieves this: – Use predicates have/1 and producerOf/2. A: if a-goal(have(milk)), bel(agent(X), producerOf(X, milk)) then send(X, !have(milk)) Koen Hindriks, Birna van Riemsdijk Multi-agent systemen Organisation • Tutorial this week: – Assignment 4: team agent for BW4T • Exam: – Specification of material to study will be posted on blackboard. – Example exams will be made available. Koen Hindriks, Birna van Riemsdijk Multi-agent systemen