Region-Based Incremental Pruning for POMDPs
This work addresses a key bottleneck in POMDP algorithms, enabling scalability for domains that were previously intractable, though it is incremental as it builds on existing incremental pruning methods.
The paper tackles the computational complexity of solving partially observable Markov decision processes (POMDPs) by introducing a region-based incremental pruning technique that divides the belief space into smaller regions for independent pruning, resulting in very significant performance gains and improved scalability to previously unmanageable domains.
We present a major improvement to the incremental pruning algorithm for solving partially observable Markov decision processes. Our technique targets the cross-sum step of the dynamic programming (DP) update, a key source of complexity in POMDP algorithms. Instead of reasoning about the whole belief space when pruning the cross-sums, our algorithm divides the belief space into smaller regions and performs independent pruning in each region. We evaluate the benefits of the new technique both analytically and experimentally, and show that it produces very significant performance gains. The results contribute to the scalability of POMDP algorithms to domains that cannot be handled by the best existing techniques.