The PELA Architecture: Integrating Planning and Learning to Improve Execution

Building architectures for autonomous rational behavior requires the integration of several AI components, such as planning, learning and execution monitoring. In most cases, the techniques used for planning and learning are tailored to the specific integrated architecture, so they could not be replaced by other equivalent techniques. Also, in order to solve tasks that require lookahead reasoning under uncertainty, these architectures need an accurate domain model to feed the planning component. But the manual definition of these models is a difficult task. In this paper, we propose an architecture that uses off-the-shelf interchangeable planning and learning components to solve tasks that require flexible planning under uncertainty. We show how a relational learning component can be applied to automatically obtain accurate probabilistic action models from executions of plans. These models can be used by any classical planner that handles metric functions, or, alternatively, by any decision theoretic planner. We also show how these components can be integrated to solve tasks continuously, under an online relational learning scheme.