Using neural networks in agent teams to speedup solution discovery for hard multi-criteria problems

Hard multi-criteria (MC) problems are computationally intractable problems requiring optimization of more than one criterion. However, the optimization of two or more criteria tends to yield not just one optimal solution, but rather a set of non-dominated solutions. As a result, the evolution of a Pareto-Optimai set of non-dominated solutions from some population of candidate solutions is often the most appropriate course of action. The non-dominated set of a population of solutions is comprised of those solutions whose criteria cannot all be dominated by those of at least one other solution in the current population. The framework we use, called the Asynchronous Team (A-Team) architecture (Talukdar, Souza, Murthy 1993), deploys teams of optimizing agents to evolve population(s) of candidate solutions to instances of hard MC problems in order to develop very good solutions. In this framework, agents embody specific heuristics geared to create, modify, or destroy any of a number of possible solutions to a problem instance. These agents are capable of choosing when and on which potential solutions they would like to work on. As a result, as the system progresses in iterations, the population of possible solutions as a whole tends to improve towards a Pareto-Optimal frontier of solutions. The Pareto Optimal frontier would consist of solutions whose individual criteria cannot be further optimized without resulting in a decline in other criteria. Currently, the method by which each agent chooses to work on particular solutions must be hand coded into the system. It can be very difficult to accomplish this since one would have to determine ahead of time which agents work best on which solutions, requiring much time and effort. In addition, the developer may introduce a ’teacher’s bias’ to the agent, hand coding incorrect decision-making based on the developer’s incorrect analysis of the agents improvement capability. Furthermore, this approach is inflexible as the hand coding done for one problem will not likely be applicable to other problems. Without hand coding this feature, agents are deployed at random with an equal likelihood. This