?
Optimal Control for Stochastic Multi-agent Systems With the Use of Parallel Hybrid Genetic Algorithm
In modern times, stochastic large-scale multi-agent systems (MAS) aimed at supporting socio-economic planning are being developed. There is a well known problem of a high computational complexity task of an optimal control for multiple agents’ behaviour in models of random interactions. In particular, agents (such as sellers and buyers) should make individualised decisions on establishing interconnections to exchange products, money or information at each moment of time. Such decisions affect the values of agents’ utility functions that, as a rule, should be maximised. In fact, each agent forms the set of individual states that define whether or not the interaction with other agents is allowed at moments of time. As a result, the dynamic programming method should be applied at the individual level of each agent maximising its own utility function, that is the extremely complex task. To overcome appropriate difficulties and seek suboptimal individual decisions in such MAS, a novel parallel hybrid real-coded genetic algorithm has been developed. The proposed hybrid method combines the use of the real-coded genetic algorithm (RCGA) for an evolutionary search, particle swarm optimisation for reducing the necessary number of model recalculations and periodically engaging ANN-based surrogate models for the fitness-function approximation. The approach allows researchers to significantly improve the time-efficiency of seeking optimal individualised decisions in MAS while keeping up their quality. Moreover, the distribution type can be one of the decision variables used in RCGAs for maximising the utility function.