Learning for Adaptive Real-time Search

TL;DR

This paper introduces an adaptive real-time search algorithm that learns heuristic functions, dynamically adjusts lookahead depth, and balances exploration and exploitation, significantly improving convergence speed in puzzle-solving tasks.

Contribution

It presents a novel algorithm that combines learning, adaptive lookahead, and user-controlled trade-offs, bridging the gap between simple greedy policies and complex planning.

Findings

Achieved 5 to 30 times faster convergence compared to classical methods.

Effectively balances exploration and exploitation during search.

Demonstrated significant improvements in sliding tile puzzle tests.

Abstract

Real-time heuristic search is a popular model of acting and learning in intelligent autonomous agents. Learning real-time search agents improve their performance over time by acquiring and refining a value function guiding the application of their actions. As computing the perfect value function is typically intractable, a heuristic approximation is acquired instead. Most studies of learning in real-time search (and reinforcement learning) assume that a simple value-function-greedy policy is used to select actions. This is in contrast to practice, where high-performance is usually attained by interleaving planning and acting via a lookahead search of a non-trivial depth. In this paper, we take a step toward bridging this gap and propose a novel algorithm that (i) learns a heuristic function to be used specifically with a lookahead-based policy, (ii) selects the lookahead depth adaptively in each state, (iii) gives the user control over the trade-off between exploration and exploitation. We extensively evaluate the algorithm in the sliding tile puzzle testbed comparing it to the classical LRTA* and the more recent weighted LRTA*, bounded LRTA*, and FALCONS. Improvements of 5 to 30 folds in convergence speed are observed.