Incorporating Energy Expenditure into Evolutionary Robotics Fitness Measures.

Abstract

Evolutionary Robotics seeks to use Evolutionary Algorithms for the purpose of creating real and simulated robots. The choice of fitness functions is a key determinant in the evolved behavior exhibited by the robot. The paper introduces energy constraints into the fitness function as a preliminary investigation into seeking to influence how evolved robots resolve tasks rather than what tasks they accomplish. An experiment is described where an artificial Sensor-Motor control system (based on the GasNet model of neuromodulated neural networks), is evolved for a robot whose task is to seek and acquire balls in a physically simulated environment. The results indicate that at least for this simple task a neural network can be evolved that achieves an energy efficient solution that is at least equal in performance to a control study where energy expenditure is not included in the fitness measure. This paper seeks to make out a case for the inclusion of energy expenditures in fitness measures, as agents are under greater selection pressure to make use of available sensory systems. It is hoped that the approach outlined in this paper will be useful in helping to develop energy efficient robots, particularly in applications such as legged locomotion.