Performance Prediction of Scalable Multi-agent Systems Using Parallel Theatre

Abstract

AbstractThis paper proposes an approach to modelling and performance prediction of large multi-agent systems, based on the theatre actor system. The approach rests on Uppaal for formal modelling, graphical reasoning and preliminary property checking, and on Java for enabling large model sizes and execution benefits on a multi-core machine. As a significant case study, the minority game (MG) binary game often used in economics, natural and social sciences is chosen for modelling and analysis. In MG, a population of agents/players compete, without explicit interactions, in the use of a shared and scarce resource. At each step, each player has to decide if to use or not the resource, and by understanding that when the majority of agents decides to exploit the resource, an inevitable congestion would arise. In classic MG, although each player learns from the experience, it is unable to improve its behaviour/performance. A genetic variant of MG is then considered which by using crossover and mutation on local strategies allows a bad-performing player to possibly improve its attitude. The paper shows an MG formal actor model, which is then transformed into Java for parallel execution. Experimental results confirm good execution speedup when the size of the model is scaled to large values, as required by practical applications.KeywordsActorsTheatre frameworkMinority gameGenetic algorithmEvolutionary learningPerformance predictionUppaalMulti-core machinesJava