Introducing Particle Swarm Optimization Into a Genetic Algorithm to Evolve Robot Controllers

Abstract :

In current robotics, growing complexity poses a challenge for classic control approaches. Evolving appropriate controllers is thus of great interest. This paper presents Genetic Behavior Graphs (Gbg), a genetic algorithm which combines elements from genetic programming and neuroevolution to develop behavior graphs network structures consisting of nodes similar to classic perceptrons, but with more functionality with respect to processing and connectivity. For local optimization of graph parameters, Gbg utilizes particle swarm optimization (Pso) intertwined with the evolution of graph structures. The algorithm's performance was evaluated on a set of black-box function approximation problems and compared to genetic programming (Gp) and the established genetic algorithm Neat, as well as with a variant of itself replacing Pso with a standard evolutionary strategy (Es). We found that the approximation of mathematically more complex functions and especifically the calculation of a robot leg's inverse kinematics is achieved significantly better by Gbg than by Gp and Neat. Also, Pso significantly improved the performance of Gbg on all problems. However, while the Gp algorithm benefited from a larger set of transfer functions of its nodes for complex problems, Gbg using Pso performed better with a smaller set.

Keywords :

Genetic Algorithms, Genetic Programming, Neuroevolution, Patricle Swarm Optimization, Robotics

Files:

gecco579_langosz.pdf