Reducing the number of evaluations required for CGDA execution through Particle Swarm Optimization methods

Abstract

Continuous Goal Directed Actions (CGDA) is a robot learning framework that encodes actions as time series of object and environment scalar features. As the execution of actions is not encoded explicitly, robot joint trajectories are computed through Evolutionary Algorithms (EA), which require a large number of evaluations. The consequence is that evaluations are performed in a simulated environment, and the optimal robot trajectory computed is then transferred to the actual robot. This paper focuses on reducing the number of evaluations required for computing an optimal robot joint trajectory. Particle Swarm Optimization (PSO) methods have been adapted to the CGDA framework to be studied and compared: naive PSO, Adaptive Fuzzy Fitness Granulation PSO (AFFG-PSO), and Fitness Inheritance PSO (FI-PSO). Experiments have been performed for two representative use cases within CGDA: the “wax” and the “painting” action. The experimental results of PSO methods are compared with those obtained with the Steady State Tournament used in the original proposal of CGDA. Conclusions extracted from these results depict a reduction of the number of required evaluations, with simultaneous tradeoff regarding the degree of fulfillment of the objective given by the optimization cost function.