![[PREV]](previous_motif.gif)
Dan L. Grecu, David C. Brown
Artificial Intelligence in Design Group
Computer Science Department
Worcester Polytechnic Institute
Abstract
Approaching design problem-solving from an AI perspective has proven the utility of multi-agent systems for this task. An agent is an autonomously acting computer program that uses knowledge to achieve goals. We use Single Function Agents (SIFAs) to explore well-delimited phenomena in parametric design. Each SIFA is a specialized problem-solver. A SiFA carries out a precise function, has a single target, and represents one point of view. For example, a selector can propose values for a spring diameter from the point of view of cost. In this paper we use selectors, praisers and critics in the parametric design of helicoidal springs. We explore a learning paradigm for these agents to make their design process more efficient.
The number of agents involved in solving a design problem is high. They need many interactions to agree on a parameter value and these values may have to be revised later due to parameter dependencies. It is hoped that through learning, agents may find better design values with less interaction. Agents that propose values - the selectors - learn by recording information about the other agents they interact with: preferences of other selectors, values considered unacceptable by critics and values praised by praisers. By repeatedly interacting with agent A, the learning agent L constructs an inductive model of the design conditions under which A would take specific design decisions. The knowledge acquired about A is later used by L in new design sessions. L will filter its proposals by anticipating responses from the agents (such as A) it has learned about.
Several experiments were carried out to investigate the benefits of the learning model. Eleven SIFAs negotiated material values for springs. Two of the SIFAs were learning agents. A reference design problem was chosen, and 21 other problems were generated by changing some of the requirements. The agents cycled through the problem set several times. The average number of interactions between agents per problem was reduced by up to 55% if the agents were allowed to learn. Learning was more efficient when the order of changes was such that a learning agent Li could first learn about the other learning agent, Lj. If Lj first had the opportunity to learn about other agents, it would reveal less of its options to Li as a result of filtering its proposals through what it has already learned. Li's task would be thus slowed down. The experiments raise the question of evaluating learning when agents learn about each other. As a result of its learned knowledge an agent starts making different proposals under the same design conditions and becomes more difficult to model.
Future research aims at exploring learning opportunities when agents negotiate on values of several parameters related through design constraints. The complexity of the task will be increased as agents with new functionalities, such as estimators and evaluators, are used. The current results also show the potential of further investigating ways to adapt what the learning process to the functional type of the agent one learns about (selector, critic etc.).