Comparing Neural Policy Learning and Neurosymbolic Search for Optimal Move Selection in Connect-4
Mete Baklacı
13/10/2026
Neural networks can learn complex patterns from data and make accurate judgments that explicit rules may not be capable of making. However, in heavily rule-based contexts such as board games, symbolic methods based on search can provide more systematic reasoning over future states compared to neural networks. Investigating the complementary capabilities of both paradigms, the present study comparatively evaluates both purely neural and neural network-guided search approaches in Connect-4 game-playing. This is done by training policy and value convolutional neural networks, where the former is used to directly generate moves whereas the latter is integrated in the minimax search algorithm as the evaluation function, forming a neurosymbolic agent. Both agents have been evaluated against symbolic game-playing benchmarks. In head-to-head evaluation, the depth-5 neurosymbolic agent defeated the policy neural network in 82.5% of games. Against the symbolic benchmark, the neurosymbolic agent achieved an 88.5% win rate compared with 69.5% for the policy neural network. The results demonstrate that neural networks can learn significant strategic information in Connect-4, while combining value learning with symbolic search can substantially improve playing performance.