Knowledge Representation for Learning How to Evaluate Partial Plans

In this paper we present some ideas for knowledge representation formalism suitable for rational agents which learn how to choose the best conditional, partial plan in any given situation. In our architecture, the agent uses an incomplete symbolic inference engine, employing Active Logic, to reason about consequences of performing actions — including information-providing ones. It utilises a simple planner to create conditional partial plans, i.e. ones which do not necessarily lead all the way to the ultimate goal. Finally, a learning module — based on ILP mechanisms — provides, from experience, knowledge on how to choose which of those plans ought to be executed. We discuss principles which should guide design of knowledge representations in order to best fit the requirements of learning process. Clearly, simply presenting all of agent’s knowledge to the ILP algorithm is very inefficient. On the other hand, for many particular applications some very effective representations are known. We compare several approaches, analysing the tradeoff between amount of domain specific knowledge provided and the quality of solutions obtained. In the experiments presented we used PROGOL for learning, and one of the conclusions of this paper is that some algorithm better suited for the particular problem of evaluating plans could significantly improve the competitiveness of domain-independent solutions