Edinburgh Napier University

  Autonomous agents commonly suffer from perceptual aliasing in which differing situations are perceived as identical by the robots sensors, yet require different courses of action. One technique for addressing this problem is to use additional internal states within a reinforcement learning system, in particular a learning classifier system. Previous research has shown that adding internal memory states can allow an animat within a cellular world to successfully navigate complex mazes. However, the technique has not previously been applied to robotic environments in which sensory data is noisy and somewhat unpredictable. We present results of using XCS with additional internal memory in the simulated Khepera environment, and show that control rules can be evolved to allow the robot to navigate a variety of problems.