Análise e comparação de desempenho do mecanismo de aprendizagem do framework de agentes MASPY

Abstract

The management of dynamic and stochastic environments requires adaptive and coordinated control systems. In this context, the integration of Reinforcement Learning with BDI (Belief-Desire-Intention) multi-agent architectures has gained prominence in the literature, as it combines symbolic goal-oriented reasoning with the flexibility of autonomous learning derived from interaction with the environment. However, there is a noticeable lack of empirical studies evaluating the practical performance and computational cost of hybrid frameworks compared to consolidated solutions. Motivated by this gap, this work aims to analyze and compare the performance of the learning mechanism of the MASPY framework, a tool developed by the LaCA-IS research group to simplify the modeling of intelligent agents. To validate its effectiveness, the study proposes solving an intelligent traffic light control problem at an urban intersection with stochastic vehicle arrivals. The methodology is based on comparing a MASPY multi-agent system, integrating a BDI architecture with Q-Learning, against a centralized baseline implemented in Python using the tabular SARSA algorithm. The performance of both approaches is evaluated through metrics such as vehicle throughput, average queue length, and resilience to demand shocks. The results demonstrate operational advantages of the MASPY-based approach regarding traffic flow capacity while also identifying the trade-offs associated with queue maintenance. It is concluded that the framework has the potential to simplify the development of complex adaptive systems, contributing to the application of hybrid architectures in dynamic control problems without compromising analytical efficiency.