Non-Blocking Batch A* (Technical Report)

TL;DR

This paper introduces Non-Blocking Batch A* (NBBA*), a heuristic search method that efficiently combines neural network heuristics with classical heuristics, reducing search expansions and improving planning speed.

Contribution

It proposes a novel non-blocking approach to batch neural network heuristic computation in A*, enabling more efficient search by integrating fast classical heuristics.

Findings

Reduces search expansions compared to blocking methods.

Performance depends on the information difference between heuristics.

Allows lazy batch computation of neural network heuristics.

Abstract

Heuristic search has traditionally relied on hand-crafted or programmatically derived heuristics. Neural networks (NNs) are newer powerful tools which can be used to learn complex mappings from states to cost-to-go heuristics. However, their slow single inference time is a large overhead that can substantially slow down planning time in optimized heuristic search implementations. Several recent works have described ways to take advantage of NN's batch computations to decrease overhead in planning, while retaining bounds on (sub)optimality. However, all these methods have used the NN heuristic in a "blocking" manner while building up their batches, and have ignored possible fast-to-compute admissible heuristics (e.g. existing classically derived heuristics) that are usually available to use. We introduce Non-Blocking Batch A* (NBBA*), a bounded suboptimal method which lazily computes the NN heuristic in batches while allowing expansions informed by a non-NN heuristic. We show how this subtle but important change can lead to substantial reductions in expansions compared to the current blocking alternative, and see that the performance is related to the information difference between the batch computed NN and fast non-NN heuristic.