A cost term in an evolutionary robotics fitness function

In order to address the domain of more complex robots and problems, robot controllers and robot structures will need to be co-evolved. To achieve the goal of the automated design of complex robotics systems and their associated control programs, a method for managing the complexity of the solution during evolution is needed. One way to do this is through the use of a cost term in the evolutionary algorithm evaluation function. In this paper we explore the use of a cost term in an evolutionary robotics evaluation function. A general form of an evaluation function containing a cost term is developed. The cost term is dynamically scaled as a function of the amount of progress the robot makes towards its goal. It has the least influence during the initial development of a solution so as not to overly restrict exploration of the search space. As the robot is able to advance further towards its goal, the cost term becomes more prominent and is used to reduce unneeded sensors and increase the efficiency of the control program. A series of experiments using a simulated robot are described. The experiments increase in difficulty and show that the general form of the evaluation function is robust and can be successfully applied in several problem domains.