Learning Predictive Models to Configure Planning Portfolios

One of the latest advances for solving classical planning problems is the development of new approaches such as portfolios of planners. In a portfolio, different base planners are run sequentially to solve a problem. Therefore, the main challenge of a portfolio planner is to define what base planners to run, in what order, and for how long. This configuration can be created manually or automatically, for instance, using machine learning techniques. In this work, a dynamic portfolio planner is described which, opposite to previous portfolio planners, is able to adapt itself to every new problem. The portfolio automatically selects the planners and the time according to predictive models. These models estimate whether a base planner will be able to solve the problem and, if so, how long it will take. The predictive models are created with machine learning techniques, using the data of the last International Planning Competition (IPC). Prediction capability of the models depend on the features extracted from the IPC results for each problem. In this work, we use a group of features extracted from the SAS+ formulation of such problems. We define different portfolio strategies, and we show that the resulting portfolios provide an improvement when compared not only with the winning planner of the last competition (LAMA), but also with less informed portfolio strategies.