Autonomous Agents and the Concept of Concepts
Paul Davidsson

This thesis has two main themes, autonomous agents and concepts. An autonomous agent is here defined as a system capable of interacting with its environment via its own sensors and effectors in order to accomplish some task. Arguments against both purely reactive and purely deliberative agent architectures are presented in favor for hybrid approaches. A novel hybrid approach based on the concept of anticipatory systems is suggested. The basic idea is to let a metalevel component run a world model faster than real time to make predictions of future states. These predictions are used to guide the agent's behavior on a high-level, whereas the low-level behavior is controlled by a reactive component. A specialization of this architecture, a linearly quasi-anticipatory agent architecture, which treats all agents in the domain (itself included) as being reactive, has been implemented. According to this approach, the state space is divided into desired and undesired regions. When the meta-level component detects that the simulated reactive system has reached an undesired state, it modifies the actual reactive system in order to avoid reaching this state. Results from both single and multiagent experiments indicate that the behavior of such agents is superior to that of the corresponding reactive agents.

Autonomous agents also provide the framework in which the representation and the acquisition of concepts are studied. However, these topics should not be studied without first answering some more fundamental questions regarding concepts, such as: What does it mean to have a concept? What functions do, or should, concepts serve? What is known about the nature of categories? Thus, by trying to answer these questions, we investigate the very concept of concepts. Although these topics seldom are discussed within Artificial Intelligence, they have received some attention in related fields, e.g., cognitive psychology and philosophy. One of the main goals of this thesis is to pull together different lines of argumentation that have emerged from the cognitive sciences in order to establish a solid foundation for further AI research. Previous approaches to concept representation are evaluated in the light of the answers to the questions above. It is concluded that none of the existing approaches is able to serve all the desired functions and that it is unrealistic to expect that any monolithic representation would be adequate. Based on this insight, a novel composite representation scheme is presented in which each component is motivated by the functions a concept should serve. Regarding the acquisition of concepts, some of the requirements that any autonomous concept learning system must meet are identified and provide the basis for an evaluation of the existing theories. A method for making any learning algorithm satisfy one such requirement, namely that of representing concepts by characteristic descriptions, is presented together with some promising experimental results. In contrast to previous methods for learning characteristic descriptions, it is possible with this method to control the degree of generalization. In addition, a new model for integrating learning by being told, learning from examples and learning by observation is outlined.

The dissertation is available at: http://www.dna.lth.se/Research/AI/Papers/PhD.ps

In addition, a limited number of hard copies are available. Send requests to paul.davidsson@dna.lth.se