Error Tolerant Learning Systems

We consider the task of a robot learning in a reactive environment by performing experiments. A reactive environment is one where changes occur in response to actions. Actors other than the learner may be present in the world. The robot performs experiments by modifying the environment and observing the outcome. These observations lead to a collection of concepts that constitute a theory of the behaviour of the environment, also called a world model. An experiment may either increase confidence in a theory or refute a theory, but it can never prove a theory. Therefore, it is possible that the robot will develop an inaccurate model of its world. This paper discusses a number of issues involved in finding and repairing faults in a world model. It also describes some preliminary results obtained from a learning program called CAP.