Learning heuristic functions for cost-based planning

In the last International Planning Competition (IPC 2011), the most efficient planners in the satisficing track were planners that used unit-cost heuristics. These heuristics ignore the real cost of the actions and return instead an estimate of the plan length to the goal. The main advantage of these heuristics compared with real-cost heuristics is that they solve a greater number of problems (also known as coverage), which has a high impact on the IPC score. However, a priori heuristics that predict the real cost should find solutions of better quality. To increase the effectiveness of real-cost heuristics and reduce the impact of their drawbacks without losing quality, we study the use of machine learning techniques to automatically obtain good combinations of those heuristics per domain. In particular, regression techniques are used to predict the real cost from any state to the goal. We use the heuristic estimations and the real costs obtained from solving easy problems as attributes. Later, we feed those instances to several machine learning techniques to obtain prediction models. All learned models approximate the real value with high correlation. Then, we implemented the most suitable model in a planner and evaluated it on harder problems. With this new planner we can solve 56 more problems than using the best real-cost heuristics for each domain separately. Our approach is also better regarding solution quality.