Designing Agent Societies Volha Bryl PhD student ICT, DIT bryl@dit.unitn.it
Agent Society + collaboration (through delegation) - competitive goals … - competitive goals Agent … Goals Capabilities Society member’s objective: use others capabilities to achieve personal goals … …
Problem Find optimal assignments: satisfy all goals of all agents or reach an compromise?
Formalization (1) Two types: Actors and Goals Goals: an actor wants a goal to be satisfied Capabilities: an actor can satisfy a goal Communication: an actor can pass (or delegate) a goal to another actor
Formalization (2) Goals can be decomposed into subgoals Only one possible decomposition Goal is satisfied iff all its subgoals are satisfied Decomposition can be performed by any actor Decomposition is actor-independent
Formalization (3) Three basic actions: an actor satisfies a goal an actor passes (or delegates) a goal to another actor an actor decomposes a goal into subgoals Solution is a list of actions that should be executed to satisfy all initial goals of all actors (or to reach an equilibrium)
Idea PLANNING Let’s look at which is in AI “the problem of finding a sequence of primitive actions to achieve some goal”
Planning problem we need to formalize actions an action has precondition and effect precondition and effect can be represented as conjunction or disjunction (for precondition only) of predicates or their negations
Planning problem. Predicates Actor’s capabilities can_satisfy (a – actor, g – goal) can_depend_on (a1 – actor, a2 – actor) Actor’s initial desires wants (a – actor, g – goal) Possible decomposition of goals can_be_decomposed (g – goal, g1 – goal, g2 – goal, …)
Planning problem. Actions 1) Satisfies (a – actor, g – goal) precondition (and can_satisfy (a, g) wants (a, g) ) effect (and done (g)* not (wants(a,g) ) * done (g – goal) is an additional predicate which means “goal g is satisfied, no one wants it anymore”
Planning problem. Actions 2) Decomposes* (a – actor, g – goal, g1 – goal, g2 – goal) precondition (and can_be_decomposed (g, g1, g2) wants (a, g) ) effect (and wants(a,g1) wants(a,g2) not (wants(a,g) ) * can be extended to more than two actions
Planning problem. Actions 3) Passes (a1 – actor, g – goal, a1 – actor) precondition (and can_depend_on (a1, a2) wants (a1, g) ) effect (and wants(a2, g) not (wants(a1, g) )
Planning problem. Intuition (1) wants are propagated through the graph and then eliminated – become done Decomposes (A, G, G1, G2) wants (A, G) wants (A, G1) wants (A, G2)
Planning problem. Intuition (2) wants are propagated through the graph and then eliminated – become done Passes (G, A, B) wants (A, G) wants (B, G) wants (A, G) Satisfies (A, G) done (G)
Example. Informal formulation I want to go to the conference We can help to get the reimbursement …to go there means: buy a ticket and get reimbursement Professor I can buy a ticket I can ask accounting office for help Accounting office
Example. PDDL formulation (define (problem example) (:domain tropos-domain) (:objects Professor Accounting_office - t_actor Go_to_conference Get_reimbursement Buy_ticket - t_goal ) (:goal (and (done Go_to_conference) ) ) (:init (can_do Accounting_office Get_reimbursement ) (can_do Professor Buy_ticket ) (can_depend_on Professor Accounting_office ) (wants Professor Go_to_conference ) (and_subgoal2 Go_to_conference Buy_ticket Get_reimbursement ) )
Example. Using the tool
Example. Using the tool
Example. Solution pass [professor, get-reimbursement, accounting-office] decomposes [professor, go-to-conference, buy-ticket, get-reimbursement] performs [professor, buy-ticket] performs [accounting-office, get-reimbursement]
Big example. Description I know English I need to write a letter to the UK embassy 1st Secretary I can use printer …to write a letter means: type, print and certify it with stamp and boss’s signature Boss 2nd Secretary I have an official stamp I can only sign letters… 3rd Secretary
Big example. Solution decomposes [boss, write-letter, send-letter, type-letter-in-english] decomposes [boss, send-letter, certify-letter, print-letter] pass [boss, type-letter-in-english, secretary-1] pass [boss, print-letter, secretary-1] performs [secretary-1, type-letter-in-english] decomposes [boss, certify-letter, stamp-letter, sign-letter] pass [boss, stamp-letter, secretary-1] performs [boss, sign-letter] pass [secretary-1, print-letter, secretary-2] performs [secretary-2, print-letter] pass [secretary-1, stamp-letter, secretary-2] pass [secretary-2, stamp-letter, secretary-3] performs [secretary-3, stamp-letter] combines [certify-letter, stamp-letter, sign-letter] combines [send-letter, certify-letter, print-letter] combines [write-letter, send-letter, type-letter-in-english]
Big example. Solution
Summing up… two views We have different views on the problem Cooperation in multi-agent communities (strategic games) Analyzing requirements These views might be though as different application areas of the proposed solution
Summing up… or what we do not have We need iterative solution building (iteration by imposing constraints on next solution) We need “partial” solution (compromise or equilibrium) and its evaluation We need the possibility of introducing new actors during the solution building