An Integrated Framework for Adaptive Reasoning About Conversation Patterns ∗ Michael Rovatsos mrovatso@inf.ed.ac.uk School of Informatics University of Edinburgh Edinburgh EH8 9LE, UK Felix Fischer fischerf@cs.tum.edu Department of Informatics Technical University of Munich 85748 Garching, Germany Gerhard Weiss weissg@cs.tum.edu Department of Informatics Technical University of Munich 85748 Garching, Germany ABSTRACT We present an integrated approach for reasoning about and learning conversation patterns in multiagent communication. The approach is based on the assumption that information about the communi- cation language and protocols available in a multiagent system is provided in the form of dialogue sequence patterns, possibly tagged with logical conditions and instance information. We describe an integrated social reasoning architecture m 2 InFFrA that is capable of (i) processing such patterns, (ii) making communication decisions in a boundedly rational way, and (iii) learning patterns and their strategic application from observation. Categories and Subject Descriptors I.2.11 [Artificial Intelligence]: Distributed Artificial Intelli- gence—Multiagent Systems, Languages and Structures General Terms Languages, Theory Keywords Agent Communication, Evolutionary Semantics, Social Reasoning, Interaction Frames 1. INTRODUCTION Compared to the long-established areas of interaction protocol and agent communication language (ACL) research [1], the de- velopment of agent architectures suitable for dealing with given communication mechanisms in practical terms has received fairly little attention. As yet, there exists no uniform framework for defin- ing the interface between the inter-agent communication layer and intra-agent reasoning, i.e. how specifications of interaction proto- cols and communication semantics influence agent rationality or, in turn, are influenced themselves by agents’ rational decision- making. ∗ A full version of this paper is available at http://homepages.inf.ed.ac.uk/mrovatso/ papers/rovatsosetal-aamas2005-full.pdf Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. AAMAS’05, July 25-29, 2005, Utrecht, Netherlands Copyright 2005 ACM 1-59593-094-9/05/0007 ...$5.00. In this paper, we attempt to tackle this problem from a very prag- matic perspective. We make very weak assumptions regarding the method used to define the available means of communication in a multiagent system (MAS), namely that it provides (i) a descrip- tion of the surface structure of communication processes (in the simplest case, traces of possible message and action sequences in agent conversations) that is tied to (ii) some form of logical con- straints (in a tractable logical language, if they are to be used in agent reasoning). In the following, we refer to such pairs of surface structure and logical constraints as conversation patterns. 2. INTERACTION FRAMES The greatest common denominator of the multitude of methods for specifying ACL semantics and interaction protocols is that they describe the surface structure of dialogues (i.e. a set of admissible message sequences) and logical constraints for the applicability of these message sequences (which may include statements about en- vironmental conditions, mental states of the participating agents, the state of commitment stores, etc.). In the most simplistic case, these structure/constraint pairs may be represented as combinations of a conversation trace and a set of logical conditions. The m 2 InF- FrA architecture [6] we describe here uses interaction frames to represent such patterns and augments them with frequency counters that allow for the definition of a probabilistic semantics. Consider the following example of such a frame: F =   5 → request(A, B, X ) 3 → do(B, X )  ,  {can(B, X )}, {can(B, pay(S)}   2 →〈[A/a], [B/b], [X /pay($100)]〉, 1 →〈[A/b], [B/a], [X /pay(S)]〉   This frame reflects the following interaction experience: A has asked B five times to perform (physical) action X , B actually did so in three of these instances. In two of the successful instances, it was a who asked and b who headed the request, and the action was to pay $100. In both cases, can(b, pay($100)) held true. In the third case, roles were swapped between a and b and the amount S remains unspecified (which does not mean that it did not have a concrete value, but that this information was abstracted away in the frame). An important feature of m 2 InFFrA frames in contrast to general conversation patterns is that they allow for storing empirical in- formation about past conversation instances that followed a certain pattern and also to distinguish between different sets of conditions that held during these enactments of a frame.