In this paper we propose a new formal model for studying reinforcement learning, based on Valiant's PAC framework. In our model the learner does not have direct access to every state of the environment. Instead, every sequence of experiments starts in a fixed initial state and the learner is provided with a "reset" operation that interrupts the current sequence of experiments and starts a new one (from the initial state). We do not require the agent to learn the optimal policy but only a good approximation of it with high probability. More precisely, we require the learner to produce a policy whose expected value from the initial state is "-close to that of the optimal policy, with probability no less than 1 \Gamma ffi . For this model, we describe an algorithm that produces such an (",ffi)-optimal policy, for any environment, in time polynomial in N , K, 1=", 1=ffi, 1=(1 \Gamma fi) and r max , where N is the number of states of the environment, K is the maximum number of actions in a...

In constructive induction (CI), the learner's problem representation is modified as a normal part of the learning process. This is useful when the initial representation is inadequate or inappropriate. In this paper, I argue that the distinction between constructive and non-constructive methods is unclear. I propose a theoretical model which allows (a) a clean distinction to be made and (b) the process of CI to be properly motivated. I also show that although constructive induction has been used almost exclusively in the context of supervised learning, there is no reason why it cannot form a part of an unsupervised regime. 1 Introduction Constructive induction (CI) is of use when the initial representation for a problem obstructs the application of ordinary inductive methods [1]. Wnek and Michalski [2] have divided constructive induction methods into several types including hypothesis-driven (HCI) methods, data-driven (DCI) methods and knowledge-driven (KCI) methods. Practical methods...

In constructive induction (CI), the learner's problem representation is modified as a normal part of the learning process. This may be necessary if the initial representation is inadequate or inappropriate. However, the distinction between constructive and non-constructive methods appears to be highly ambiguous. Several conventional definitions of the process of constructive induction appear to include all conceivable learning processes. In this paper I argue that the process of constructive learning should be identified with that of relational learning (i.e., I suggest that what constructive learners really learn is relationships) and I describe some of the possible benefits that might be obtained as a result of adopting this definition.

The paper considers the quantification of inductive bias in concept learning. It argues that some of the best known measures of bias (eg.