Time-dependent genetic algorithm and its application to quadruped's locomotion

Abstract

Genetic algorithms (GAs) are widely used in machine learning and optimization. This paper proposes a time-dependent genetic algorithm (TDGA) based on real-coded genetic algorithm (RCGA) to improve the convergence performance of functions over time such as a foot trajectory. TDGA has several distinguishing features when compared with traditional RCGA. First, individuals are arranged over time, and then the individuals are optimized in sequence. Second, search spaces of design variables are newly comprised of processes of reductions for search spaces. Third, the search space for crossover operations is expanded to avoid local minima traps that can occur in new search spaces up to the previous search space before performing any reduction of search space, and boundary mutation operation is performed to the new search spaces. Computer simulations are implemented to verify the convergence performance of the robot locomotion optimized by TDGA. Then, TDGA optimizes the desired feet trajectories of quadruped robots that climb up a slope and the impedance parameters of admittance control so that quadruped robots can trot stably over irregular terrains. Simulation results clearly represent that the convergence performance is improved by TDGA, which also shows that TDGA could be broadly used in robot locomotion research.