This paper is about representations for Artificial Intelligence systems. All of the results described in it involve engineering the representation to make AI systems more effective. The main AI techniques studied here are varieties of search: path-finding in graphs, and probablilistic searching via simulated annealing and genetic algorithms. The main results are empirical findings about the granularity of representation in implementations of genetic algorithms. We conclude by proposing a new algorithm, called “Long-Term Evolution,” which is a genetic algorithm running on (...) an evolving problem description. We see this as modelling the evolution of a species from simpler (more coarsely described— fewer genes) types of organisms to more complex ones. The results, which are reported here of our experiments with the algorithm make it seem a promising optimisation technique. (shrink)

Negotiation between two individuals is a common task that typically involves two goals: maximize individual outcomes and obtain an agreement. However, research on the simplest negotiation tasks demonstrates that although naive subjects can be induced to improve their performance, they are often no more likely to achieve fully optimal solutions. The present study tested the prediction that a decrease in a particular type of argumentative behavior, substantiation, would result in an increase in optimal agreements. As substantiation behaviors depend primarily on (...) supplied content of the negotiation task, it was also predicted that substantiation behavior would be reduced by curtailing the content. A 2 × 2 experimental design was employed, where both negotiation tactics (list of tactics present versus absent) and negotiation task content (high versus low) were varied to determine the processes leading beyond solution improvement to solution optimality. Sixty‐one dyads engaged in a two‐party, four‐issue negotiation task. All negotiations were videotaped and analyzed. Although the list of negotiation tactics resulted in improved performance, only the content manipulation resulted in a significant increase in dyads achieving optimal solutions. Analyses of the coded protocols indicated that the key difference in achieving optimality was a reduction in persistent substantiation‐related operators (substantiation, along with single‐issue preferences and procedures) and an increase in a complex macro‐operator, multi‐issue offers that reduced the problem space, facilitating the search for optimality. (shrink)