Learning context conditions for BDI plan selection

An important drawback to the popular Belief, Desire, and Intentions (BDI) paradigm is that such systems include no element of learning from experience. In particular, the so-called context conditions of plans, on which the whole model relies for plan selection, are restricted to be boolean formulas that are to be specified at design/implementation time. To address these limitations, we propose a novel BDI programming framework that, by suitably modeling context conditions as decision trees, allows agents to learn the probability of success for plans based on previous execution experiences. By using a probabilistic plan selection function, the agents can balance exploration and exploitation of their plans. We develop and empirically investigate two extreme approaches to learning the new context conditions and show that both can be advantageous in certain situations. Finally, we propose a generalization of the probabilistic plan selection function that yields a middle-ground between the two extreme approaches, and which we thus argue is the most flexible and simple approach.