Learning and Joint Deliberation through Argumentation in MAS Santi Ontañón (GeorgiaTech) Enric Plaza (IIIA-CSIC) dijous 12 de novembre de 2009 1
Outline • Introduction • Justified Predictions in MAS • Arguments and Counterexamples • Argumentation-based MAL • Experimental Evaluation • Conclusions & Future Work dijous 12 de novembre de 2009 2
Committee dijous 12 de novembre de 2009 3
Committee Input Deliberation Aggregation Output Committee : ( 1 ) A group of people o ffi cially delegated ( elected or appointed ) to perform a function, such as investigating, considering, reporting, or acting on a matter. Team : A number of persons associated in some joint action. A group organized to work together. Coalition : A combination or alliance, esp. a temporary one between persons, factions, states, etc. An alliance for combined action, especially a temporary alliance of political parties. dijous 12 de novembre de 2009 3
Ensemble Effect Joint Joint Prediction prediction Ensemble better than effect best P i Aggregation (e.g. Voting) Predictors P 1 P 2 ... P n-1 P n Bagging / Boosting / other... Data Set dijous 12 de novembre de 2009 4
Committee of agents Joint Joint Prediction Ensemble prediction effect better than best A i Voting A 1 A 2 ... A n-1 A n Agents DS DS 1 DS 2 DS n n-1 dijous 12 de novembre de 2009 5
Deliberation + Voting A 1 A n Agents A 2 A n-1 ... Deliberation (Argumentation framework) dijous 12 de novembre de 2009 6
Deliberation + Voting Joint Prediction Agreement Voting A 1 A n Agents A 2 A n-1 ... Deliberation (Argumentation framework) dijous 12 de novembre de 2009 6
Introduction • CBR agents • solve problems and learn from them • How could CBR agents collaborate? • solving and/or learning from cases • Argumentation process • to mediate agent collaboration • achieving a joint prediction dijous 12 de novembre de 2009 7
Learning vs Cooperating • Why to learn in problem solving? • to improve accuracy, range, etc • Why to cooperate in problem solving? • to improve accuracy, range, etc • Learning-cooperation continuum • learning agents that cooperate by arguing dijous 12 de novembre de 2009 8
CBR cycle Problem Retrieve New New Case Retrieved Case Retrieved Case Case n Reuse i Precedent a Case t e R Domain Knowledge Revised Solved Case Case Revise dijous 12 de novembre de 2009 9
Argumentation in Multi-Agent Learning • An argumentation framework for learning agents • Justified Predictions as arguments • Individual policies for agents to • generate arguments and • generate counterarguments • select counterexamples dijous 12 de novembre de 2009 10
Justified Prediction Justification : A symbolic description with the information relevant to determine a specific prediction Case 1 Case 2 Traffic_light : red Traffic_light : green Cars_passing : no Cars_passing : no Solution: wait Problem Solution: wait Solution: cross Traffic_light : red Justification Cars_passing : no Case 3 Case 4 Traffic_light : red Traffic_light : red Traffic_light : green Cars_passing : yes Cars_passing : yes Solution: wait Solution: wait Retrieved cases dijous 12 de novembre de 2009 11
Justification example Solution: hadromerida Justification: D 1 Spikulate Skeleton Sponge Megascleres Megascleres P LID Spikulate Smooth form : tylostyle Uniform length : no New skeleton sponge External External features features Gemmules : no Case Base of A 1 ! ! " ! ! ! " " #" $%&'()*'+&%" , " " dijous 12 de novembre de 2009 12
Justification example Solution: hadromerida Justification: D 1 Spikulate Skeleton Sponge Megascleres Megascleres P LID Spikulate Smooth form : tylostyle Uniform length : no New skeleton sponge External External features features Gemmules : no Case Base of A 1 ! ! " ! ! ! " " #" $%&'()*'+&%" , " " The predicted solution is hadromerida because the smooth form of the megascleres of the spiculate skeleton of the sponge is of type tylostyle , the spikulate skeleton of the sponge has not uniform length , and there are no gemmules in the external features of the sponge. dijous 12 de novembre de 2009 12
Counterargument generation Solution: astrophorida ! ! " ! ! ! " " #" $%&'()*'+&%" , " " Justification: D 2 Spikulate Skeleton Sponge Megascleres Megascleres Spikulate Smooth form : tylostyle LID Uniform length : no skeleton External External features features Gemmules : no Growing Grow : massive Growing : Case Base of A 2 " " # ! ! ! # " #" %-.'(/$('+&%" , # " dijous 12 de novembre de 2009 13
Argument types • Justified Prediction : An argument α α = � A i , P , + , D � endorsing a individual prediction • Counterargument : An argument β β = � A 2 , P , − , D 2 � offered in opposition to an argument α • Counterexample : A case c c = � P 1 , −� contradicting an argument α dijous 12 de novembre de 2009 14
Case-based Confidence ! ! ! ! ! " #" " " $ " + ! # " ! . ! / ! !"#"# ! 0 % 1 - + + + - + + - - + + - #$%& '$%& () $*&+, - ! dijous 12 de novembre de 2009 15
Preference on Arguments α = � A i , P , + , D � + Confidence on an argument based on - + + + - ? cases Y Y N C ( α ) = Y + N dijous 12 de novembre de 2009 16
Preference on Arguments(2) α C ( α ) = 4 + 5 = 0 . 8 C ( β ) = 2 β 3 = 0 . 66 - + + + - + - ? P dijous 12 de novembre de 2009 17
Preference on Arguments(2) α C ( α ) = 4 Preferred + 5 = 0 . 8 C ( β ) = 2 β 3 = 0 . 66 - + + + - + - ? P dijous 12 de novembre de 2009 17
Preference on Arguments(2) α C ( α ) = 4 Preferred + 5 = 0 . 8 C ( β ) = 2 β 3 = 0 . 66 - + + + - + - Joint Confidence ? P Y A 1 + Y A 2 + 1 α α C ( α ) = Y A 1 + Y A 2 + N A 1 + N A 2 + 2 α α α α dijous 12 de novembre de 2009 17
Relations between arguments a) b) c) α α α + + + β β β - + - ? ? ? P P P α = � A 1 , P , + , D 1 � α = � A 1 , P , + , D 1 � α = � A 1 , P , + , D 1 � β = � A 2 , P , − , D 2 � β = � A 2 , P , + , D 2 � β = � A 2 , P , − , D 2 � dijous 12 de novembre de 2009 18
Relations between arguments a) b) c) α α α + + + β β β - + - ? ? ? P P P α = � A 1 , P , + , D 1 � α = � A 1 , P , + , D 1 � α = � A 1 , P , + , D 1 � β = � A 2 , P , − , D 2 � β = � A 2 , P , + , D 2 � β = � A 2 , P , − , D 2 � Incomparable Consistent Counterargument dijous 12 de novembre de 2009 18
Relations between cases and justified predictions α = � A i , P , + , D � a) b) α = � A i , P , + , D � c) α = � A i , P , + , D � + + + + - - ? ? ? c = � P 1 , + � c = � P 1 , −� c = � P 1 , −� P P P dijous 12 de novembre de 2009 19
Relations between cases and justified predictions α = � A i , P , + , D � a) b) α = � A i , P , + , D � c) α = � A i , P , + , D � + + + + - - ? ? ? c = � P 1 , + � c = � P 1 , −� c = � P 1 , −� P P P Endorsing case Irrelevant case Counterexample dijous 12 de novembre de 2009 19
Argument Generation • Generation of a Justified Prediction • LID generates a description α . D subsuming P • Generation of a Counterargument • if no Counterargument can be generated then: • Selection of a Counterexample dijous 12 de novembre de 2009 20
Counterargument Generation α + • Counterarguments are generated based β - on the specificity criterion • LID generates a description β . D ? subsuming P and subsumed by α . D P dijous 12 de novembre de 2009 21
Selection of a Counterexample α = � A i , P , + , D � + • Select a case c subsumed by α . D and endorsing a different solution class. - ? c = � P 1 , −� P dijous 12 de novembre de 2009 22
AMAL protocol Assertions of n agents at round t H t = � α t 1 , ..., α t n � Justified prediction asserted assert ( α ) in the next round Agent states a counterargument ß rebut ( β, α ) contradict ( α t i ) = { α ∈ H t | α.S � = α t i .S } Set of contradicting arguments for agent A i at round t (those predicting a different solution) dijous 12 de novembre de 2009 23
Agents assert α DELIBERATION at Round t & Agent owning the token Joint t All Agree? YES Solution NO Nobody Agent+token Voting new args Generates Generates CE CA Better CA Rebut CA Better CA Agent'+CE NO than ? to Agent' than ? α � α YES NO YES Agent Agent' New arg asserts CA asserts CA dijous 12 de novembre de 2009 24
Argument Generation contradict ( α t i ) = { α ∈ H t | α .S � = α t i .S } Generate CA for each May not found a CA for each β 1 ... β k If empty generates CE Select argument with a generated CA that has lowest confidence α i Most likely to "convince" the Select CA for that other agent to change assertion argument β i dijous 12 de novembre de 2009 25
Recommend
More recommend
