Adaptive path planning: algorithm and analysis

摘要

We present a learning algorithm that improves path planning by using past experience to enhance future performance. The algorithm relies on an existing path planner to provide solutions to difficult tasks. From these solutions, an evolving sparse network of useful robot configurations is learned to support faster planning. More generally, the algorithm provides a framework in which a slow but effective planner may be improved both cost-wise and capability-wise by a faster but less effective planner coupled with experience. We analyze the algorithm by formalizing the concept of improvability and deriving conditions under which a planner can be improved within the framework. The analysis is based on two stochastic models, one pessimistic (on task complexity), the other randomized (on experience utility). Using these models, we derive quantitative bounds to predict the learning behavior. We use these estimation tools to characterize the situations in which the algorithm is useful and to provide bounds on the training time. In particular, we show how to predict the maximum achievable speedup.