Cache-Efficient Dynamic Programming MDP Solver

Automated planning research often focuses on developing new algorithms to improve the computational performance of planners, but effective implementation can also play a significant role. Hardware features such as memory hierarchy can yield substantial running time improvements when optimized. In this paper, we propose two state-reordering techniques for the Topological Value Iteration (TVI) algorithm. Our first technique organizes states in memory so that those belonging to the same Strongly Connected Component (SCC) are contiguous, while our second technique optimizes state value propagation by reordering states within each SCC. We analyze existing planning algorithms with respect to their cache efficiency and describe domain characteristics which can provide an advantage to each of them. Empirical results show that, in many instances, our new algorithms, called eTVI and eiTVI, run several times faster than traditional VI, TVI, LRTDP and ILAO* techniques.

Proceedings of the 26th European Conference on Artificial Intelligence