NeuroSymbolic-RLNet: a neuro-symbolic reinforcement learning framework for transparent sequential decision-making
Abstract
The integration of symbolic reasoning with deep reinforcement learning presents a promising paradigm for achieving transparent and interpretable sequential decision-making in complex environments. This work introduces NeuroSymbolic-RLNet, a novel hybrid framework that combines symbolic state transition graphs with neural embeddings to enable explainable policy learning while maintaining superior performance in stochastic tasks. The proposed architecture employs a dual-layer reasoning mechanism where symbolic rules guide high-level decision strategies, while neural networks handle low-level feature extraction and policy optimization. Comprehensive evaluation across robotics control benchmarks, financial decision scenarios, and adversarial environments demonstrates that NeuroSymbolic-RLNet achieves 23.7% improvement in sample efficiency, 31.2% better interpretability scores, and 18.9% enhanced robustness compared to traditional deep reinforcement learning methods. The framework successfully addresses critical limitations in current RL approaches including lack of transparency, poor generalization to unseen scenarios, and vulnerability to adversarial perturbations. Extensive ablation studies and real-world deployment experiments validate the effectiveness of our neuro-symbolic integration approach across urban, rural, and mixed operational environments.
