Orthogonal Evolution of Teams: Evolving Teams of Specialized Members for Robust Intelligence
This project will develop and validate a novel cooperative, co-evolutionary algorithm design for multi-agent systems: Orthogonal Evolution of Teams (OET). Many real-world problems are too large and too complex to expect a single, monolithic intelligent agent to solve them successfully. Monolithic agents tend to overlook specialized sub-domains within a larger problem space and thus to make errors on those sub-domains. As problems grow progressively larger and more complex, this weakness will become increasingly critical. Thus, considerable research has focused on team or ensemble approaches, generating agents that consist of integrated subsystems each of which is specialized to find solutions within a sub-domain of the total problem space. For a team approach to succeed it must meet two criteria: The team members must be relatively successful and the team members must collaborate in a way that improves the performance of the team as a whole.
Research has shown that evolutionary techniques are effective and broadly applicable to the evolution of teams. However, existing evolutionary approaches have significant weaknesses. They either produce relatively poor team members, or they do not sufficiently encourage specialization between the team members. The specific aim of this research is to investigate OET's ability to meet the two criteria and to validate its performance as compared to other team approaches on a number of problems, including artificial problems specifically designed to test features of its performance, typical test problems from the evolutionary computation and classification fields, and the real-world problem of identifying disordered regions in proteins. Validation will include using the predicted fault (or error) rate model to measure the specialization of the team members with OET and with other team based learning algorithms, comparison of results with OET to benchmark algorithms and published results, and entering an OET based algorithm in the biannual Critical Assessment of Techniques for Protein Structure Prediction (CASP) experiment.
The results will lead to significantly improved algorithms for generating teams of classifiers, predictors, and similar learning agents with wide applications in a number of fields. More directly, the work will create better identifiers of disordered proteins, a significant problem in its own right. The work will support, directly and indirectly, a number of graduate student research programs, and it will be incorporated into existing undergraduate programs.