Constraint Facilitated Learning in Artificial Systems

The goal of our study was to determine whether the implementation of sensory and motor constraints on a developing robot would significantly improve learning progress. Developing robots were initiated in a rectangular hallway facing a wall whose color was dependent on the robot’s movement and position. Throughout the developmental process, the robot had to predict future sensorimotor vectors based on past experiences and report the error in this prediction. Experiments were run on both constrained and unconstrained robots. A comparison was made between the prediction successes in the final stages of development in both the constrained and unconstrained robots. We found no significant evidence showing that the addition of constraints led to lower prediction error. We also found no significant evidence of decrease in prediction error in the unconstrained agents. The data collected led us to believe that neither the constrained nor unconstrained experimental groups were successful in learning to predict their environment and that the designed world was too difficult for the robot to learn with the given parameters and developmental architecture.