with R einforcemen T ) Learning Agents : A Preliminary Report

We present a framework for building agents that lea rn using SMART, a system that combines stochastic model acquisition with reinforcement lea rning to enable an agent to model its environment through experience and subsequently for m action selection policies using the acquired model. We extend an existing algorithm for automati c creation of stochastic strips operators (Oates et. al 1995) as a preliminary method of environment modelling. We then define the process of generation of future states using these operators a nd an initial state and finally show the process by which the agent can use the generated states to for m a policy with a standard reinforcement learning algorithm. The potential of SMART is exem plified using the well-known predator prey scenario. Results of applying SMART to this environ ment and directions for future work are discussed.