VRAIL (Vectorized Reward-based Attribution for Interpretable Learning) is a bi-level framework for value-based reinforcement learning (RL) that learns interpretable weight representations from state features. VRAIL consists of two phases: a deep learning (DL) phase that fits an estimated value function using state features, and a RL phase that shapes the learning through a potential-based reward transformation. The estimator can be modeled in linear or quadratic form to attribute importance to individual features and their interactions. Experimental results on the Taxi-v3 environment demonstrate that VRAIL improves training stability and convergence compared to standard DQN without any environment modification. Further analysis highlights the ability of VRAIL to discover semantically meaningful subgoals, such as passenger possession, to generate human-interpretable actions. The results suggest that VRAIL serves as a general, model-independent reward-shaping framework that enhances both learning and interpretability.
