Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Argumentation in Artificial Intelligence: Theoretical Foundations and Technological Applications De donde venimos… Carlos Chesñevar & Guillermo Simari Desde 1995 funciona en la UNS el Departamento de Cs. e Ing. en Ciencias de la Computación, y dentro de éste, el Laboratorio de Investigación y Desarrollo en Inteligencia Artificial (LIDIA) CONICET and Laboratory of R&D in A.I. Department of Computer Science and Engineering Universidad Nacional del Sur Bahía Blanca, Argentina Univ. Nacional del Sur (Bahía Blanca, Argentina) http://lidia.cs.uns.edu.ar A modo de introducción…. ¿Qué pretende este curso? • La argumentación puede verse de manera abstracta como el intercambio de razones a favor y en contra de cierta conclusión. • Brindar una introducción a los conceptos teóricos fundamentales asociados a la argumentación en Inteligencia Artificial. • En los últimos años, la argumentación ha ganado creciente importancia en el ámbito de la Inteligencia Artificial (IA) como vehículo para facilitar una interacción racional. • Introducir algunos de los sistemas lógicos utilizados para argumentación, particularmente la denominada Programación en Lógica Rebatible (DeLP) y sus extensiones. • En un contexto de múltiples agentes, la argumentación provee herramientas para diseñar, diseñar modelar e implementar formas de interacción entre agentes. • Mostrar cómo la argumentación puede aplicarse en el contexto multiagente para modelar toma de decisiones, decisiones negociación, negociación etc. etc • • En un contexto de un único agente, puede utilizarse argumentación para modelar su razonamiento, teniendo en cuenta preferencias, dinamicidad del entorno, etc. Ilustrar aplicaciones y sistemas inteligentes argumentación desarrollados recientemente. • Objetivo final: que quien realice el curso pueda acceder a literatura científica vinculada a la argumentación en IA, aplicarla en su ámbito de investigación, y/o profundizar en aquellos temas que le hayan resultado de interés. • Este curso apunta a dar una introducción a la argumentación en el contexto de agentes inteligentes. Argumentation in Artificial Intelligence, 2008 3 basados en Argumentation in Artificial Intelligence, 2008 4 Sobre el material a utilizar Outline - Part 1 • En este curso trabajaremos con transparencias que resumen los principales contenidos del curso. • Introduction. Motivations for this course. Goals. • Default Reasoning: problems, alternatives. Argumentation as a way of formalizing default reasoning. • Las transparencias se complementan con distintos artículos en revistas y trabajos en congresos recientes. • Systems for Defeasible Argumentation. Elements. Argument, Counterargument, defeat, dialectical analysis. Unique status vs. Multiple status approach. • Las transparencias estarán en castellano o inglés. • Conclusions Argumentation in Artificial Intelligence, 2008 5 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 6 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Main references – Part 1 Outline H. Prakken, G. Vreeswijk. Logical Systems for Defeasible Argumentation, in D. Gabbay (Ed.), Handbook of Philosophical Logic, 2nd Edition, 2002. • (Very brief) Introduction to Multiagent Systems • C.Chesñevar, A.Maguitman, R.Loui. Logical Models of Argument. In ACM Computing Surveys, Argument Surveys Dec. Dec 2000. 2000 • Conclusions • Trevor J.M. Bench-Capon, Paul E. Dunne. Argumentation in Artificial Intelligence. Artificial Intelligence Journal, Volume 171, Issues 10-15, JulyOctober 2007, Pages 619-641. • Argumentation in Artificial Intelligence, 2008 • What is argumentation? Fundamentals 7 Overview Argumentation in Artificial Intelligence, 2008 Ubiquity, Interconnection, Intelligence ¨ Five ongoing trends have marked the history of computing: ¨ As processing capability spreads, sophistication (and intelligence of a sort) becomes ubiquitous. ¨ What could benefit from having a processor embedded in it…? • ubiquity; • interconnection; • intelligence; • delegation; and • human-orientation ¨ Internet is powerful…Some researchers are putting forward theoretical models that portray computing as primarily a process of interaction. ¨ The complexity of tasks that we are capable of automating and delegating to computers has grown steadily. Credits: some of these slides are based on Michael Wooldridge’s lecture notes for his book “An Introduction to MAS” (Wiley & Sons, 2002) Argumentation in Artificial Intelligence, 2008 9 Argumentation in Artificial Intelligence, 2008 Delegation, Human-Orientation 10 Programming progression… ¨Programming has progressed through: ¨ Computers are doing more for us – without our intervention. Next on the agenda: fly-by-wire cars, intelligent braking systems… • machine code; • assembly language; ¨ Programmers conceptualize and implement software in t terms off higher-level hi h l l – more human-oriented h i t d– abstractions. • machine-independent programming languages; • sub-routines; b ti • procedures & functions; ¨ The movement away from machine-oriented views of programming toward concepts and metaphors that more closely reflect the way we ourselves understand the world. • abstract data types; • objects; Argumentation in Artificial Intelligence, 2008 8 to agents. 11 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 12 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Where does it bring us? Interconnection and Distribution ¨ Delegation and Intelligence imply the need to build computer systems that can act effectively on our behalf. ¨ Interconnection and Distribution have become core motifs in Computer Science. ¨ But Interconnection and Distribution, coupled y to represent p our best with the need for systems interests, implies systems that can cooperate and reach agreements (or even compete) with other systems that have different interests (much as we do with other people). ¨ This implies: • The ability of computer systems to act independently. • The ability of computer systems to act in a way that represents our best interests while interacting with other humans or systems. Argumentation in Artificial Intelligence, 2008 13 So Computer Science expands… ¨ A multiagent system is one that consists of a number of agents, which interact with oneanother. ¨ All of these trends have led to the emergence of a new field in Computer Science: Multiagent Systems. ¨ In the most general case, agents will be acting on behalf of users with different goals and motivations. ¨ An agent is a computer system that is capable of independent action on behalf of its user or owner (figuring out what needs to be done to satisfy design objectives, rather than constantly being told). ¨ To successfully interact, they will require the ability to cooperate, coordinate, and negotiate with each other, much as people do. 15 Multiagent Systems Argumentation in Artificial Intelligence, 2008 Agents as Technologies ¨ In Multiagent Systems, we address questions such as: • How can cooperation emerge in societies of selfinterested agents? • What kinds of languages can agents use to communicate? i t ? • How can self-interested agents recognize conflict, and how can they (nevertheless) reach agreement? • How can autonomous agents coordinate their activities so as to cooperatively achieve goals? Argumentation in Artificial Intelligence, 2008 14 Multiagent Systems: a Definition ¨ These issues were not studied in Computer Science until recently. Argumentation in Artificial Intelligence, 2008 Argumentation in Artificial Intelligence, 2008 17 From: Agent Technology: A Roadmap for Agent-based Computing. M. Luck et. Al, 2005, Agentlink. Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 16 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Multiagent Systems Multi-agent systems ¨ An agent needs the ability to make internal reasoning: • • • • • ¨ Agents need to: • exchange information and explanations • resolve conflicts of opinions • resolve conflicts of interests • make joint decisions Reasoning about beliefs, desires, … Handling inconsistencies Making decisions Generating, revising, and selecting goals ... they need to engage in dialogues ¨ Argumentation provides a powerful metaphor for solving these problems! Argumentation in Artificial Intelligence, 2008 19 Argumentation in Artificial Intelligence, 2008 20 The role of argumentation Why to study argumentation in MAS? ¨ For internal reasoning of single agents: ¨ Argumentation plays a key role for achieving the goals of the above dialogue types • Reasoning about beliefs, goals, ... • Making decisions • Generating, revising, and selecting goals ¨ Argument = Reason for some conclusion (belief, action, goal, etc.) ¨ For F interaction i t ti b between t multiple lti l agents: t ¨ Argumentation = Reasoning about arguments → decide on conclusion • Exchanging information and explanations • Resolving conflicts of opinions ¨ Dialectical argumentation = Multi-party argumentation through dialogue • Resolving conflicts of interests • Making joint decisions Argumentation in Artificial Intelligence, 2008 21 Argumentation in Artificial Intelligence, 2008 The role of argumentation Types of Dialogues Argumentation plays a key role for reaching agreements: Type ¨ Additional information can be exchanged ¨ The opinion of the agent is explicitly explained (e.g., arguments t in i favor f off opinions i i or offers, ff arguments t in i favor of a rejection or an acceptance) ¨ Agents can modify/revise their beliefs / preferences / goals 23 Initial Situation Main Goal Participant’s aims Subtypes Information Personal Spreading Gain, pass on, seeking ignorance knowledge show, or hide knowledge • Expert consultation • Interview • Interrogation Persuasion Conflicting Resolution Persuade the Beliefs of conflict other(s) by verbal means • Dispute Inquiry ¨ To influence the behavior of an agent (threats, rewards) Argumentation in Artificial Intelligence, 2008 22 General Growth of Find a proof or Ignorance knowledge destroy one & agreemt’ Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 • Scientific Research • Investigation 24 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari A negotiation dialogue Types of Dialogues Buyer: Can’t you give me this Peugeot 407 a bit cheaper? Initial Situation Type Main Goal Participant’s aims Reach a Influence decision Outcome Deliberation Need for action Making Conflict of interests & a deal need for cooperation Negotiation Get the best for oneself Subtypes Seller: Sorry that’s the best I can do. Why don’t you go for a Peugeot 307 instead? • Board meeting • War planning Buyer: I have a big family and I need a big car (B1) Seller: Modern 307’s are becoming very spacious and would easily fit in a big family. (S1) • Bargaining • Union negotiation • Land dispute Buyer: I didn’t know that, let’s also look at 307’s then. Typology by Walton & Krabbe, 1995 Argumentation in Artificial Intelligence, 2008 25 Argumentation in Artificial Intelligence, 2008 26 Module for sensors and control of effectors Touch Smell Taste Generic Agent Language Sensors receive perceptions Environ nment Module “What to do n next?” Sensors Hearing Sight Emotions ¿What to do? Cognitive Behavior Muscles Effectors execute those chosen actions to be carried out… Thought Effectors Memory Do-it-yourself agent Artificial Intelligence: A Modern Approach, 2nd Ed., S.Russell & P.Norvig 2003 Architecture Sensors Beliefs Plans Sensors Beliefs Plans Desires Interpreter Desires Intentions Intentions Argumentation-Based Reasoning Engine! Argumentation! Actuator Actuator Actuator Actuator Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Simple Reactive Agent Reactive Agent with Inner State Agent Agent Sensors How the World Evolves Rules Condition-action What to do? Effectors Agent with Explicit Goals Utility-based agent Goals Observations About the current State of The world Which are the conseq. of doing action A Agent Sensors State Observations About the current state Of The world How the World Evolves Consequences from actions How good is the state I would achieve? Utility What to do now? Which are the conseq ences consequences Of doing action A? Environ nment Consequences from Actions Sensors Environ nment How The World evolves What to do? Effectors Argumentation-Based Reasoning Engine! State Observations About the world Consequences of the actions Argumentation-Based Reasoning Engine! Agent Sensors Environ nment Rules Condition-action Environ nment Observations about the world State What to do now? Effectores Argumentation-Based Reasoning Engine! Argumentation-Based Reasoning Engine! Effectors Outline • (Very brief) Introduction to Multiagent Systems • What is argumentation? Fundamentals “Argumentation is a verbal and social activity of reason aimed at increasing (or decreasing) the acceptability of a controversial standpoint for the listener or reader, by putting forward a constellation of propositions intended to justify (or refute) the standpoint before a rational judge.” (van Eemeren et al., 1996) Argumentation in Artificial Intelligence, 2008 35 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 36 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Elements of Argument Argumentation: two flavors Argumentation ¨ Premises: Logic-oriented g view Linguistic-oriented g view • If Iraq has weapons of mass destruction (WMDs), then it poses a threat to world peace • Iraq has WMDs ¨ Conclusion: Borrows many concepts from legal reasoning, rhethorics, etc. Follows a more “traditional” style in symbolic AI… Cross fertilization • Has enabled the development of automated argumentation systems (e.g. DeLP). Gained importance in the context of the World Wide Web (Semantic Web tools) Argumentation in Artificial Intelligence, 2008 Credits for some of the next slides: Iyad Rahwan (COMMA 2006 and AAAI 2007 presentations) and Henry Prakken (EAAAS 2007). 37 Types of Argument Iraq poses a threat to world peace. Argumentation in Artificial Intelligence, 2008 38 Toulmin’s Schema (1950) [Walton 2006] ¨ Deductive argument • If premises are true, the conclusion is true • E.g. mathematical proof D So, Q, C ¨ Inductive argument Since W • Generalisation from observation • True to some degree of confidence Unless, R On account of B ¨ Presumptive argument • Conclusion is plausible, given the premises • Argument is defeasible as it may be overridden Argumentation in Artificial Intelligence, 2008 D for Data Q for Qualifier C for Claim 39 W for Warrant B for Backing R for Rebuttal ¨ Claim: Iraq poses a threat to world peace ¨ Data: Iraq has WMDs ¨ Warrant: Countries with WMDs threaten peace ¨ Backing: 80% of countries with WMDs since 1920 disrupted world peace ¨ Rebuttal: Unless these countries are democracies ¨ Qualifier: Presumably, certainly, possibly, etc. Argumentation in Artificial Intelligence, 2008 40 Critical Questions Walton’s Argument Schemes Critical Questions allow to undermine an argument scheme. ¨ Argument Schemes classify typical forms of premises and conclusions ¨ E.g. Scheme for Argument from Expert Opinion ¨ E.g. Scheme for Argument from Expert Opinion • Premise: Source E is an expert in the subject domain S • Premise: E asserts that proposition A in domain S is true • Conclusion: A may plausibly be taken to be true ( Premise: Source E is an expert in the subject domain S ( Premise: E asserts that proposition A in domain S is true ( Conclusion: A may plausibly be taken to be true ¨ Critical Questions: ¨ E.g. Actual Argument from Expert Opinion • Premise: The CIA says that Iraq has WMDs • Premise: The CIA are experts on WMDs • Conclusion: Iraq has WMDs ( ( ( ( ( ( ¨ In natural language, some premises may be implicit. Argumentation in Artificial Intelligence, 2008 41 Expertise: How credible is expert E? Backup Evidence: Is A supported by evidence? Field: Is E an expert in the field that A is in? Opinion: Does E's testimony imply A? Trustworthiness: Is E reliable? Consistency: Is A consistent other experts’ testimony? Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 42 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Rebutting via Critical Question: Presumptions vs. Exceptions Using Argument Schemes ¨ There are many other schemes, e.g. ¨ Undermine a presumption, e.g. ( Argument from analogy ( Argument from negative consequence ( … ( Backup Evidence: Is A supported by evidence? ( You presume that the CIA have evidence for Iraq’s WMDs, but this is not true. ¨ Important role in argumentative analysis ¨ Show that an exception holds, e.g. ( ( ( ( ( Consistency: Is A consistent other experts’ testimony? ( Actually, many other experts disagree that Iraq has WMDs Argumentation in Artificial Intelligence, 2008 43 Analyse text text, speech speech, etc etc. Identify argument scheme used Make implicit premises explicit Use critical questions for evaluation Argumentation in Artificial Intelligence, 2008 44 Argument schemes: general form Combining Arguments ¨ The same as logical inference rules Premise 1, … , Premise n Therefore (presumably), conclusion ¨ Critical questions provide “pointers” which make an argument schema unapplicable (e.g. undermining presumptions, showing that an exception holds, etc.) Argumentation in Artificial Intelligence, 2008 45 Argumentation in Artificial Intelligence, 2008 46 Argument Schema: “Normative syllogism” Argument Schema: Statistical syllogism ¨ P and if P then as a rule Q ¨ P and if P then usually Q is a reason to believe that Q is a reason to accept that Q ¨ Critical question: are there exceptions? John Pollock 3 Birds usually fly ( How does a lawyer argue for exceptions to a rule? ¨ Critical question: subproperty defeater? 3 Say legislation makes an exception ( Conflicting generalisation about an exceptional class 3 Say it is motivated by the rule’s purpose 3 Penguins don’t fly 3 Find an overruling principle 3 Argue that rule application has bad consequences Argumentation in Artificial Intelligence, 2008 47 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 48 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Argument Schema: Witness testimony Argument Schema: Temporal persistence P is true at T1 and T2 > T1 Witness W says P Therefore (presumably), P is still true at T2 Therefore (presumably), P ¨ Critical questions: ¨ Critical q questions: • Is W sincere? (veracity) • Was P known to be false between T1 and T2? • Did W really see P? (objectivity) • Is the gap between T1 and T2 too long? • Did P occur? (observational sensitivity) Argumentation in Artificial Intelligence, 2008 49 Argument Schema: Arguments from consequences Argumentation in Artificial Intelligence, 2008 50 Types of Dialgoue (Walton and Krabbe, 1995) Action A brings about good consequences Dialogue Type Dialogue Goal Initial situation Persuasion resolution of conflict conflict of opinion Negotiation making a deal conflict of interest Deliberation reaching a decision need for action Information seeking exchange of information personal ignorance Inquiry growth of knowledge general ignorance Therefore (presumably), A should be done ¨ Critical questions: • Does A also have bad consequences? • Are there other ways to bring about the good consequences? ( We will come back to this later … ) Argumentation in Artificial Intelligence, 2008 51 Argumentation in Artificial Intelligence, 2008 52 The main problem: representing and reasoning with commonsense knowledge Defeasible reasoning in AI Suppose we want to represent knowledge of the real world ¨ Reasoning is generally defeasible • Assumptions, exceptions, uncertainty, ... 1. Birds fly. 2. Penguins do not fly. 3. Penguins are birds. 4. Tweety is a bird. 5. Tweety is a penguin. ¨ AI formalises such reasoning with non-monotonic logics • Default logic, logic etc … • New premisses can invalidate old conclusions K= ¨ Argumentation logics formalise defeasible reasoning as construction and comparison of arguments Bird(x) Æ fly(x) Penguin(x) Æ ¬fly(x) Penguin(x) Æ Bird(x) Bird(tweety). Penguin(tweety). First order logic cannot deal with inconsistent information ! K A fly(tweety) K A ¬fly(tweety) This is a central problem in Artificial Intelligence… how to deal with defaults? Argumentation in Artificial Intelligence, 2008 53 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 54 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Some issues in default reasoning How to represent and reason with defaults? Typical problems in (non-monotonic) default reasoning: 1) Representation of defaults: e.g. Birds usually fly 1) Representation of defaults: e.g. Birds usually fly Solutions: 2) Inconsistency handling: identify relevant subsets of consistent information. a) Abnormality conditions: an additional predicate “ab” is included. ∀x Bird(x) ∧ ¬ab1(x) → Canfly(x) y gp preferred models 3)) Identifying Many approaches have been developed: E.g.: ∀x Bird(x) ∧ ¬BrokenWing(x) → Canfly(x) • Default logic (Reiter, 1980) • Preferred subtheories (Brewka, 1989) Problem: how many different types of abnormalities? • Circunscription (McCarthy, 1987) Æ Qualification problem • Others… Argumentation in Artificial Intelligence, 2008 55 How to represent and reason with defaults? Argumentation in Artificial Intelligence, 2008 56 How to represent and reason with defaults? Solutions: Solutions: b) Default logic (Reiter, 1980) c) Preferred Subtheories (Brewka, 1989) Bird(x) : Canfly(x) / Canfly(x) Defaults can be also formalized as formulas in first-order logic (ordinary implications). Defeasibility is captured in terms of consistency handling strategies. Some subtheories are preferred over others. “If it is provable that X is a bird, and it is not provable that X cannot fly, then we may infer that X can fly” D={ Bird(x) : Canfly(x) / Canfly(x) } (1) Bird Æ canfly (2) Penguin Æ ¬canfly (3) bird (4) penguin W = { Bird(Tweety), ∀x. Penguin(x) Æ ¬ Canfly(x) } Then D ∪ W A Canfly(tweety) But D ∪ W ∪ { Penguin(tweety) } G Canfly(tweety) Argumentation in Artificial Intelligence, 2008 How to represent and reason with defaults? 58 Argumentation systems (AS) are “yet another way” to formalize common-sense reasoning. Non-monotonicity arises from the fact that new premises may give rise to stronger counterarguments, which in turn will defeat the original argument. d) Defeasible Rules Defaults can be also formalized in terms of special rules called “defeasible rules”, used to represent tentative knowledge. Argumentation in Artificial Intelligence, 2008 Argumentation in Artificial Intelligence, 2008 Systems for defeasible argumentation. Generalities Solutions: Bird(X) canfly(X) {(2),(4)} Two conflicting subtheories 57 ∀x Bird(X) Æ canfly(X) {(1),(3)} 1) Normality condition view: an argument = standard proof from a set of premises + normality statements. g is an attack on such a normality y A counterargument statement. (If X is a bird, then X can fly) 2) Inconsistency handling view: an argument = standard proof from a consistent subset of the premises. A counterargument is an attack on a premise of an argument. (Birds usually can fly) 3) Semantic view: constructing ‘invalid’ arguments (wrt the semantics) is allowed in the proof theory. A counterargument is an attack on the use of an inference rule which deviates from a preferred model. 59 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 Views on default reasoning from an argumentation perspective 60 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Systems for Defeasible Argumentation Argumentation process (Amgoud) According to Prakken & Vreeswijk (2002), there are five common elements to systems for defeasible argumentation: Constructing arguments Defining the interactions between arguments Definition of Underlying Logical Language Evaluating the strengths of arguments Definition of Argument Defining the status of arguments Definition of Conflict among Arguments Definition of Defeat among Arguments Drawing conclusions using a consequence relation Comparing decisions using a given principle Definition of Status of Arguments Inference problem Argumentation in Artificial Intelligence, 2008 61 The underlying logic: Arguments & Logical consequence ¨ Argumentation Systems are constructed starting from a logical language and an associated notion of logical consequence for that language. Decision making problem Argumentation in Artificial Intelligence, 2008 62 Systems for Defeasible Argumentation Definition of Underlying Logical Language Definition of Argument ¨ The logical consequence relation helps to define what will be considered an argument. argument D fi i i off C Definition Conflict fli among A Arguments ¨ This consequence relation is monotonic, i.e., new information cannot invalidate arguments as such, but rather give rise to counterarguments. Definition of Defeat among Arguments Definition of Status of Arguments ¨ Arguments are seen as proofs in the chosen logic. Argumentation in Artificial Intelligence, 2008 63 Language for Argumentation Argumentation in Artificial Intelligence, 2008 64 Systems for Defeasible Argumentation The object language is basically a logical language: Definition of Underlying Logical Language ¨ Constants (tweety, opus, etc.) and variables (X,Y,Z). Definition of Argument ¨ Predicate symbols (flies (flies, bird, bird etc.). etc ) D fi i i off C Definition Conflict fli among A Arguments ¨ Connectives (∧, →, ¬, not) Definition of Defeat among Arguments ¨ Special symbol for defeasible rule: e.g. or ⇒ Definition of Status of Arguments ¨ Negation is important! Argumentation in Artificial Intelligence, 2008 65 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 66 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Argument as a ‘proof’ The concept of argument Arguments are presented under different forms: K = { bird(tweety), penguin(tweety), bird(X)Æ flies(X), penguin(X)Æ bird(X), penguin(X) Æ ¬flies(X)} ¨ An inference tree grounded in premises. ¨ A deduction sequence. Using modus ponens, from K we can infer flies(tweety) and ¬ flies(tweety) ¨ A pair (Premises (Premises, Conclusion), Conclusion) leaving unspecified the particular proof, in the underlying logic, that leads from the Premises to the Conclusion. Logic does not “care” in the contents of proofs, Logic is static. Argumentation does care, as it is a dynamic process… ¨ A completely unspecified structure, such as in Dung’s abstract framework for argumentation (1995). flies(tweety) ¬flies(tweety) bird(tweety) penguin(tweety) penguin(tweety) Argumentation in Artificial Intelligence, 2008 67 Argumentation in Artificial Intelligence, 2008 The concept of argument Types of arguments: (Kraus et al. 98, Amgoud & Prade 05) ¨ Explanations (involve only beliefs) • Kstrict = { bird(tweety), penguin(tweety), bird(X)Æ flies(X), penguin(X)Æ bird(X), penguin(X) Æ ¬flies(X)} bird(tweety) Tweety flies because it is a bird ¨ Threats (involve beliefs + goals) Kdefeasible = {bird(X)Æ flies(X), penguin(X) Æ ¬flies(X)} ¬flies(tweety) 68 Different kinds of arguments... That’s why it is important to distinguish tentative knowledge from strict knowledge… flies(tweety) Two conflicting “arguments” derived from K Two conflicting “arguments” derived from Kstrict ∪ Kdefeasible • You should do α otherwise I will do β • You should not do α otherwise I will do β ¨ Rewards (involve beliefs + goals) penguin(tweety) penguin(tweety) • If you do α, I will do β • If you don’t do α, I will do β … Argumentation in Artificial Intelligence, 2008 69 Systems for Defeasible Argumentation 70 Conflict, Attack, Counterargument The notion of conflict (Counterargument or Attack) between arguments is typically discussed discriminating three cases: Definition of Underlying Logical Language ¨ Rebutting attacks: arguments with contradictory conclusions. conclusions Definition of Argument D fi i i off C Definition Conflict fli among A Arguments ¨ Assumption attack: attacking non-provability assumptions. Definition of Defeat among Arguments Definition of Status of Arguments Argumentation in Artificial Intelligence, 2008 Argumentation in Artificial Intelligence, 2008 ¨ Undercutting attacks: an argument that undermines some intermediate step (inference rule) of another argument. 71 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 72 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Rebutting and assumption attacks Rebutting is symmetric, e.g.: ‘Tweety flies because it is a bird’ versus Tweety doesn’t fly because it is a penguin’. tweety flies Undercutting attack ¨ An argument challenges the connection between the premises and the conclusion. Assumption attack: Tweety flies because it is a bird and it is not provable that Tweety is a penguin’ versus Tweety is a penguin’ ¬tweety flies tweety flies ¬⎡ p,q,r / h ⎤ h penguin tweety p q r Tweety flies because all the birds I’ve seen fly not(penguin tweety) Argumentation in Artificial Intelligence, 2008 73 Direct vs. Indirect Attack ¬p Argumentation in Artificial Intelligence, 2008 74 Systems for Defeasible Argumentation These types of attack could be direct and indirect. Direct attack Definition of Underlying Logical Language Definition of Argument Indirect attack ¬p p I’ve seen Opus; it is a bird and it doesn’t fly D fi i i off C Definition Conflict fli among A Arguments s Definition of Defeat among Arguments p Argumentation in Artificial Intelligence, 2008 Definition of Status of Arguments 75 Argumentation in Artificial Intelligence, 2008 76 Defeat: Comparing Arguments Defeat: Comparing Arguments ¨ The notion of conflict does not embody any form of comparison; this is another element of argumentation ¨ Argumentation systems vary in their grounds for evaluation of arguments. One common criterion is the specificity principle, which prefers arguments based on the most specific defaults. ¨ Defeat has the form of a binary relation between arguments, standing for • ‘attacking ‘ tt ki and d nott weaker’ k ’ (d defeat f t) • ‘attacking and stronger’ defeats (strict defeat) flies(opus) ¨ Terminology varies: ‘defeat’ (Simari, 1989; Prakken & Sartor, 1997), ‘attack’ (Dung, 1995; Bondarenko et. al 1997) and ‘interference’ (Loui, 1998). Argumentation in Artificial Intelligence, 2008 bird(opus) 〈 A, flies(opus) 〉 77 ¬flies(opus) ≤ bird(opus), broken_wing(opus) 〈 B, ¬flies(opus) 〉 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 78 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Defeat among arguments via specificity If something moves when touched, it I usually not dead If something looks dead, it is usually dead If a spider is dead, usually it is not dangerous If something is a spider, it is usually dangerous Black widow (bw) is a spider Black widow (bw) looks dead. Black widow (bw) moves when touched. ¨ Specificity criterion: prefers arguments (a) with more information or (b) more direct dangerous(bw) h p h p Argument A for dangerous (bw) spider(bw) < A B spider(bw), dead(bw) ~dead (bw) Discussion: “Computing Generalized Specificity” (F.Stolzenburg, A.García, C.Chesñevar , G.Simari). Journal of Applied Non-Classical Logics, Vol.13, No. 1, pp.87-113, 2003. Argumentation in Artificial Intelligence, 2008 moves_when_touched (bw) < A, dangerous(bw) > < B, ~dangerous(bw) > looks_dead(bw) A B Yes. Argument B for ~dangerous(bw) defeats A (it is strictly more specific than A, as it is more informed) ~dangerous (bw) < ?- dangerous(bw). Argument C for ~dead(bw) defeats B (it attacks argument B at an inner subargument, being as specific as the subargument being attacked). < C, ~dead(bw) > 79 80 Defeat: Comparing Arguments Defeat: comparing arguments ¨ However, it has been argued that specificity is not a general principle of commonsense reasoning, but rather a standard that might (or might not) be used. ¨ In Simari&Loui’s framework, specificity is used as a default, but it is ‘modular’: any other preference relation defined among arguments could be used. ¨ Some researchers even claim that general, domainindependent principles of defeat do not exist exist, or are very weak. ¨ In Dung’s, defeat is an abstract notion, left undefined undefined. ¨ In Bondarenko’s framework, defeat is limited to attack between arguments (there is no preference at all!) ¨ Some even argue that the evaluation criteria are part of the domain theory, and should also be debatable. ¨ Other comparison criteria are possible… What do you think? Argumentation in Artificial Intelligence, 2008 81 Defeat: comparing arguments Argumentation in Artificial Intelligence, 2008 82 Systems for Defeasible Argumentation ¨ Defeat is basically a binary relation on a set of args. ¨ But ... it just tells us something about two arguments, not about a dispute (that may involve many args.) Definition of Underlying Logical Language ¨ A common situation is reinstatement as in the example below (where an argument C reinstates an argument A b d by defeating f ti argumentt B) Definition of Argument D fi i i off C Definition Conflict fli among A Arguments Definition of Defeat among Arguments Definition of Status of Arguments A Argumentation in Artificial Intelligence, 2008 B† C 83 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 84 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Status of Arguments Status of arguments ¨ The last element in our ontology comes into play... the definition of Status of Arguments. ¨ Status of arguments can be computed either in ‘declarative’ or ‘procedural’ form. ¨ This notion is the actual output of most Arg.Sys and arguments are divided into (at least) two classes: ¨ In the declarative form usually requires fixed-point definitions, and establishes certain sets of arguments as acceptable (in the context of a set of premises and a evaluation criteria) but without defining a procedure for testing whether a given argument is a member of this set. • Arguments with which a dispute can be ‘won’ won • Arguments with which a dispute can be ‘lost’ • Arguments that leave the dispute ‘undecided’ ¨ ‘Procedural form’ amounts to defining such a procedure for acceptability. ¨ Usual terminology: ‘justified’ or ‘warranted’ vs. ‘defeated’ or ‘overruled’ vs. ‘defensible’, etc. Argumentation in Artificial Intelligence, 2008 85 Status of arguments Declarative form of argumentation Argumentation in Artificial Intelligence, 2008 86 Argument-based Semantics ¨ Which conditions on sets of arguments should be satisfied? Procedural form of argumentation ¨ We will assume as background • A set Args of arguments defeat defined over it. it • A binary relation of ‘defeat’ Def. 1: Arguments are either justified or not justified Argumentationtheoretic semantics 1. An argument is justified if all arguments defeating it (if any) are not justified. Proof Theory 2. An argument is not justified if it is defeated by an argument that is justified. Argumentation in Artificial Intelligence, 2008 87 Argumentation in Artificial Intelligence, 2008 Argument-based Semantics 88 Example: Even cycle A =“Nixon was a pacifist because he was a quaker” B =“Nixon wasn’t a pacifist because he was a republican” Example: Consider three arguments A, B and C There are two status assignment that satisfy Def 1 † A B Def. 1: Arguments are either justified or not justified A B C 1. An argument is justified if all arguments defeating it (if any) are not justified. Argument A and C are justified; argument B is not. Argumentation in Artificial Intelligence, 2008 2. An argument is not justified if it is defeated by an argument that is justified. 89 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 90 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Argument-based Semantics The Unique-Status-Assignment Approach In the literature, two approaches to the solution of this problem can be found. This idea could be presented in two different ways: ¨ First approach: changing Def. 1 in such a way that there is always precisely one possible way to assign a status to arguments. Undecided conflicts get the status ‘not not justified’ justified . ¨ Using a fixed-point operator Allowing unique-status assignment (u.s.a). ¨ Second approach: allowing multiple assignments, defining an argument as ‘genuinely’ justified iff it is justified in all possible assignments. ¨ Given a recursive definition of justified argument Allowing multiple-status assignment (m.s.a). Argumentation in Artificial Intelligence, 2008 91 Argumentation in Artificial Intelligence, 2008 92 Problems with Unique-Status Assignment Problems with Unique-Status Assignment There are some problems when evaluating unique-status assignment. Example: Floating Arguments / Floating Conclusions There are some problems when evaluating unique-status assignment. Example: Floating Arguments / Floating Conclusions p D C A A - = Juan es mendocino porque vive en Mendoza B- = Juan es porteño por la tonada que tiene A = Juan es argentino porque es mendocino. B = Juan es argentino porque es porteño. But since C is defeated by A and B, we could argue that C “floats” on A and B, and therefore it should be overruled. And then D is justified. AA B Argumentation in Artificial Intelligence, 2008 93 Argumentation in Artificial Intelligence, 2008 94 Using Multiple-Status Assignment Problems with Unique-Status Assignment ¨ A second way to deal with competing arguments of equal strenght is to let them induce two alternative status assignments. There are some problems when evaluating unique-status assignment. Example: Floating Arguments / Floating Conclusions ¨ Evaluating outcomes from alternative status assignments let us determine when an argument is justified. D p The unique-status approach is inherently unable to capture floating arguments and conclusions. C A BB B Argumentation in Artificial Intelligence, 2008 AA Def. : (Status assignment) Given a set S of args ordered by a binary defeat relation, an status assignment sa(S) is a function which maps every argument in S into {in,out}, such that: i. BB A is in iff all args defeating it (if any) are out. ii. A is out if it is defeated by an arg that is in. 95 Argumentation in Artificial Intelligence, 2008 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 96 Argumentación en Inteligencia Artificial Carlos Iván Chesñevar – Guillermo Ricardo Simari Example Classifying Arguments Def. : Given a set S of arguments ordered by a binary defeat relation, an argument A is – justified iff it is ‘in’ in all sa(S). B A B A ( ) – overruled iff it is ‘out’ in all sa(S) – defensible iff it is ‘out’ in some sa(S), ‘in’ in others. Def. : (Justification) Given a set S of arguments ordered by a binary defeat relation, an argument is justified iff it is in in all possible status assignments to S. Argumentation in Artificial Intelligence, 2008 ¨ Are the two approaches are equivalent? ¨ The answer is no. 97 ¨ Some researchers say that the difference between the two approaches can be compared with the ‘skeptical’ vs. ‘credulous’ attitude towards drawing defeasible conclusions ... C A ¨ m.s.a is more convenient for identifying sets of arguments that are compatible with each other. B ¨ u.s.a considers arguments on an individual basis. ¨ What are the status assignments? ¨ There are no status assignments! Argumentation in Artificial Intelligence, 2008 98 Comparing the two approaches Problems with Multiple-Status Assignment A Argumentation in Artificial Intelligence, 2008 99 Argumentation in Artificial Intelligence, 2008 END OF PART 1 101 Escuela Internacional CACIC 2008 Chilecito, Argentina – 06 al 10 de octubre de 2008 100