Document Type

Restricted

Advisor

Gary Parker

Publication Date

2026

Comments

This paper is restricted to the Connecticut College campus until December 8, 2026.

Abstract

This thesis investigates whether neuroevolution can produce heuristic functions for A* search on Sokoban that are drastically more parameter-efficient than a conventional gradient-trained baseline. Sokoban is a combinatorial planning puzzle that is both NPhard and PSPACE-complete, and solving it at scale requires an accurate cost-to-go heuristic to guide search. The standard approach fixes a neural network architecture and trains it by gradient descent; this thesis investigates whether evolutionary search over network topology can find heuristics that are competitive in accuracy while using orders of magnitude fewer parameters.

Four heuristic approaches are implemented and compared on the medium-difficulty split of the Boxoban benchmark, a dataset of procedurally generated 10×10 Sokoban levels released by Google. A parameter-free Manhattan-distance heuristic, summing per-box minimum distance to a goal, anchors the minimum baseline. A CNN cost-to-go network adapted from DeepCubeA, with roughly 2 million parameters trained by AdamW on cost-to-go labels derived from Breadth-First Search solution trajectories, anchors the upper-limit baseline. NEAT (NeuroEvolution of Augmenting Topologies) evolves both the topology and the weights of a feed-forward heuristic from scratch using the same labels, growing complexity only where it improves prediction accuracy. CoDeepNEAT coevolves populations of network modules and blueprint topologies, combining evolutionary architecture search with per-genome gradient-descent weight training.

The headline finding is that neuroevolved heuristics fall on opposite sides of the Manhattan minimum baseline: pure NEAT, despite searching over both topology and weights, fails to clear it, while CoDeepNEAT clears it using roughly 7.7× fewer parameters than the CNN upper-limit baseline. CoDeepNEAT is therefore the only neuroevolved method in this thesis that beats the parameter-free analytical baseline; the addition of Adam gradient descent on each assembled genome is what closes the gap that pure NEAT cannot. The results show that neuroevolution can beat a parameter-free analytical baseline on Boxoban, but only when evolutionary architecture search is paired with gradient-based weight optimization on inputs of this dimensionality.

Download

Share

COinS
 

The views expressed in this paper are solely those of the author.