Using Inverse Resolution to Learn Relations from Experiments

We are concerned with learning relations in a reactive environment. A learning agent observes sequences of actions that may change the properties or relationships of objects in the world. The observed sequence is then used to form a theory which can be generalised and tested by imitation. That is, the learner attempts to perform its own sequence of actions. If the test completes as expected then the generalisation is accepted. However, if it fails then a regeneralisation occurs. That is, further generalisations are found to explain the difference between expectation and reality. Inverse Resolution is the primary generalisation mechanism. However, learning by experimentation in a reactive environment causes difficulties that inverse resolution alone cannot handle. We describe the additions required to the theory that enable us to deal with ‘surprise’ in an experiment, i.e. learning from unexpected results. New inverse resolution operations are used to allow us to generalise from partial matches between our theory and the world, thus enabling us to explain what went wrong with a theory. 1 Address all correspondence to Claude Sammut (email: [email protected])