170 likes | 327 Vues
This presentation explores the innovative approach of using random walks to improve local search strategies in planning tasks. It discusses the limitations of traditional random walks and introduces Random Walk-Driven Local Search (RW-LS) as a solution to escape local minima and plateaus. The research presents results from IPC-2011 benchmarks, showcasing how RW-LS generates better solutions despite some speed trade-offs. Future work aims to develop this method into a global search algorithm and integrate multi-heuristic techniques, further advancing planning capabilities.
E N D
Planning via Random Walk-Driven Local Search Fan Xie HootanNakhost Martin Müller Presented by: HootanNakhost
Outline • Random Walks Planning • Problems of Random Walks • Random Walk-Driven Local Search • Experiments • Conclusion and Future Works
Arvand Planner[Nakhostand Müller 2009] • Exploration using random walks to overcome the problem of local minima and plateaus. • Jumping greedily exploits the knowledge gained by the random walks.
Aras: Plan Improving Postprocessor[Nakhostet al. 2011] • Expand a neighbor search space along the input plan. • Output the shortest path in the neighbor search space
Problems of Random Walks • Fails in Narrow Exit Path Search Space (we will explain later) • Poor Plan Quality
Random Walk-Driven Local Search • Local Greedy Best First search (local GBFS) • Perform one random walk from the node going to be expanded • Keep the best random walk (lowest h) • Jump to the best state (in GBFS or end-point of random walk)
Random Walk-Driven Local Search For every node in the open list, it has two heuristic value: hn: the heuristic value of the node itself hr: the heuristic value of the end-point of the random walk starting from the node Nodes in the open list are ordered by a linear combination of hn and hr(W = 100 in our experiments): W * hn+ hr
Analysis of RW-LS • Advantage: • A small local search can help escape some small Narrow-Exit-Path-Search-Space • Generally, generates better solutions • Disadvantage: • Slow down speed
Larger Problems Easy domains with scalable generators are scaled to get larger problems: The number in parentheses are the max number used in the IPC-2011 benchmark
Conclusion and Future Work • Contribution: • RW-LS: A strong algorithm combining local search and random walks, implemented in the Planner Arvand-LS. • Motivated and Developed larger problems • Future work: • Make it a global search algorithm • Make it one entry in a portfolio planner • Add multi-heuristic into the algorithm