Learning from Previous Execution to Improve Route Planning
|Title||Learning from Previous Execution to Improve Route Planning|
|Publication Type||Conference Proceedings|
|Year of Conference||2013|
|Authors||Gohde J, Boddy M, Shackleton H|
|Conference Name||ICAPS Workshop on Planning and Learning|
In this paper, we describe a specific approach to iterative planning in the domain of off-road route planning, in which the objective is to find a cost-minimal path from one point to another. In iterative planning we are concerned with finding a way to solve a succession of planning problems, improving the system’s behavior over time.1 For example, this improvement might come about through improved heuristics, leading to more effective search of the space of possible plans, or through corrections or additions to the domain model used in planning. In this work, we take the latter approach, modifying the domain model based on differences between plans generated using the existing model and “good” plans.
We have implemented our approach to iterative planning for generating off-road routes in a system called G2I2. In Section 2, we briefly discuss the route planning problem. Section 3 presents the current implementation of G2I2. Section 4 describes the learning model. The rest of the paper presents a set of experiments undertaken and summarizes the results obtained (Section 5), and discusses the implications of those results and the relationship of this work to other approaches to learning planning models (Section 7). Finally, we offer some concluding discussion in Section 7.