Controlling evaluation duration in On-line, on-board evolutionary robotics

Abstract

In this paper, we evaluate parameter control techniques for on-line and on-board evolutionary robotics. The devised approach augments an algorithm for on-line controller adaptation ((μ+1) ON-LINE) with a scheme for dynamic control of the evaluation time (τmax). We measure the performance of the approach in experiments that combine Fast-Forward and Temperature-Driven Fast-Forward tasks. The results with preselected optimal static evolution time are compared to those where τmax is dynamically controlled using a number of different control schemes. The experiments show that the devised approaches for parameter control can improve the performance of robots as the controller adapts to changes in the environment or task objective. A dynamic τmax-selection also eliminates the need to tune this parameter prior to deployment, letting the evolutionary process control the evaluation time of each robot controller depending on the current task and environment.