Using Genetic Algorithms to Capture Behavioral Traits Exhibited by Knowledge Based Robot Agents

One of the fundamental issues in the field of Evolutionary Robotics (ER) is that of selection and formulation of an appropriate performance training metric. In recent years, proof-of-concept studies have shown that simple behavioral robotics problems, such as homing and foraging, are amenable to ER methods. However, the question of scalability remains unresolved. Several researchers have shown that straightforward ER methods fail to produce viable results on more complex problems if the problem is not partitioned or preprocessed in some fashion before applying ER methods. In many cases, the knowledge required to preprocess the problem is equivalent to that which would be needed to formulate a purely rule knowledge-based controller, hence, in such cases, ER methods provide no real benefit. In this work, we present research results of an investigation into the feasibility of using observed behavior in an environment to train artificial neural network-based robotic controllers to function in that same environment. Robot agents were allowed to navigate through a selection of artificial life simulation environments under the influence of knowledge-based controllers. At each time-step, the simulated robot sensor inputs and actuator outputs were recorded. The resulting input and output data were used to train artificial neural network based controllers for the different environments. The resulting neural network based controllers were then used to control robots in similar environments and were found to exhibit features of the original knowledge-based controllers. Key Words--Evolutionary Robotics, Behavioral Robotics, Evolutionary Neural Networks, Artificial Life