Catálogo de publicaciones - libros

SeaSpeak is “English for maritime communications.” It is a restricted, specially-designed dialect of English used in merchant shipping and accepted as an international standard. This paper discusses, in the context of SeaSpeak, two key problems in the formalization of any such restricted, specially-designed language, viz., representing the illocutionary force structure of the messages, and formalization of such reference-fixing devices from ordinary language as pointing and use of demonstratives. The paper conducts the analysis in terms of Kimbrough’s FLBC agent communication language.

This paper presents a fully formal integration of Jones’s logical theory of speech acts as signalling conventions with Kimbrough’s Formal Language for Business Communication (FLBC). The work is part of a larger programme of logicism in the context of electronic commerce. Speech acts are an especially apt subject for this programme because of their pervasiveness and importance in communication for all commerce, electronic or not. The paper demonstrates that the conventionist view of speech acts, embodied in Jones’s logical theory, fits naturally with Kimbrough’s FLBC and with the Basic Messaging Framework for business communications. Further, the paper provides an illustration of how the resulting integrated theory might be implemented in practice through logic programming.

The author provides an in-depth look at a moderately complex conversation as represented by a finite state machine, a representation used by an established agent communication system. He compares this to a statechart-based method used for Moore’s conversation policy framework and observes that these methods differ in their level of detail, the usefulness of parts of the graphical representation, and in the grouping of events. The remainder of the paper demonstrates how a multi-agent conversation policy can be used to control the flow of messages, contrasts this with how messages are handled via an inference-based process, and shows how the inference-based processing can be integrated with the policy-based handling in order to deal with exceptions to the policy.

In this paper we describe a research project centering on experiments in which game-playing evolving agents are used to investigate the value of information. Specifically, in these experiments we define populations of agents whose strategies evolve towards those that have better restocking strategies for their supply chain. The agents evolve their strategies in order to minimize costs (either for themselves or for their value chain). We describe several different experiments in which we will vary the abilities of agents both to gather and to store more information. Part of the results of this project will be related to the value of information and computational capabilities: Is it always better to have more information? If not, what are the conditions under which less information is better? The culminating experiment is one in which evolving agents compete to sell information to other evolving agents playing their roles in a supply chain.

This paper discusses the principal reasons for, and prospective opportunities of, simulating financial markets using an architecture based on artificial agents. The paper then discusses in detail the design and architecture of a simulator for financial markets. The Gaia methodology was employed in the development of MAFiMSi (Multi-Agent Finanacial Market Simulator), a general-purpose finacial market simulator of a dealer-type market. MAFiMSi is implemented as a library of C++ classes that currently support a stand-alone market simulation.

Coalition formation games form an important subclass of mixed-motive strategic situations, in which players must negotiate competitively to secure contracts. This paper compares the performance of two learning mechanisms, reinforcement learning and counterfactual reasoning, for modeling play in such games. Previous work [CK04] found that while the former type of agent converged to theoretical solutions, they did so much more slowly than human subjects. The present work addresses this issue by allowing agents to update extensively based on counterfactual reasoning.

Automated negotiation by artificial adaptive agents (AAAs) holds great promise for electronic commerce, but non-trivial, practical issues remain. Published studies of AAA learning of negotiation strategies have been based on artificial environments that include complete payoff information for both sides of the bargaining table. This is not realistic in applied contexts. Without loss of generality, we consider the case of a seller who knows its own preferences over negotiation outcomes but will have limited information about the private values of each customer. We propose a learning environment that takes advantage of partial information likely to be available to the vendor. General strategies are learned for a group of similar customers - a market segment - through a simulation approach and a genetic learning algorithm. In addition, we systematically further relax constraints on the opponent’s preferences to further explore AAA learning in incomplete information environments.

We report on a series of computational experiments with artificial agents learning in the context of games. Two kinds of learning are investigated: (1) a simple form of associative learning, called Q-learning, which occurs in state space, and (2) a simple form of learning, which we introduce here, that occurs in policy space. We compare the two methods on a number of repeated 2×2 games. We conclude that learning in policy space is an effective and promising method for learning in games.

We examine learning by artificial agents in repeated play of Cournot duopoly games. Our learning model is simple and cognitively realistic. The model departs from standard reinforcement learning models, as applied to agents in games, in that it credits the agent with a form of conceptual ascent, whereby the agent is able to learn from a consideration set of strategies spanning more than one period of play. The resulting behavior is markedly different from behavior predicted by classical economics for the single-shot (unrepeated) Cournot duopoly game. In repeated play under our learning regime, agents are able to arrive at a tacit form of collusion and set production levels near to those for a monopolist. We note that Cournot duopoly games are reasonable approximations for many real-world arrangements, including hourly spot markets for electricity.

An important dimension of system and mechanism design, working memory, has been paid insufficient attention by scholars. Existing literature reports mixed findings on the effects of the amount of working memory on system efficiency. In this note, we investigate this relationship with a computational approach. We design an intelligent agent system in which three agents, one buyer and two bidders, play an Exchange Game repeatedly. The buyer agent decides whether to list a request for proposal, while the bidders bid for it independently. Only one bidder can win on a given round of play. Once the winning bidder is chosen by the buyer, a transaction takes place. The two parties of the trade can either cooperate or defect at this point. The decisions are made simultaneously and the payoffs essentially follow the Prisoner’s Dilemma game. We find that the relationship between working memory and the efficiency of the system has an inverted U-shape, i.e., there seems to be an optimal memory size. When we mixed agents with different memory sizes together, agents with the same amount of working memory generate the most efficient outcome in terms of total payoffs.