Explainable AI-aided Feature Selection and Model Reduction for DRL-based V2X Resource Allocation

TL;DR

The paper tackles the lack of explainability in DRL-driven V2X resource allocation by introducing a model-agnostic, two-stage SHAP-based XAI framework that ranks state features and reduces input dimensionality. It employs a centralized-training, decentralized-execution MADRL algorithm to jointly optimize sub-band allocation and transmit power under URLLC reliability constraints, achieving near-maximum sum-rate with significantly reduced model complexity. The authors demonstrate that masking low-importance features via SHAP can preserve around 97% of the original MADRL performance while cutting average training time by up to ~26% and reducing trainable parameters by up to ~46% in networks with eight vehicular pairs. This approach enhances practical deployment by improving interpretability, reducing communication overhead for weight updates, and enabling scalable operation in dynamic vehicular environments. Future work includes adapting the XAI framework to other wireless resource allocation problems and integrating it with digital twin and meta-learning paradigms to further improve adaptability.

Abstract

Artificial intelligence (AI) is expected to significantly enhance radio resource management (RRM) in sixth-generation (6G) networks. However, the lack of explainability in complex deep learning (DL) models poses a challenge for practical implementation. This paper proposes a novel explainable AI (XAI)- based framework for feature selection and model complexity reduction in a model-agnostic manner. Applied to a multi-agent deep reinforcement learning (MADRL) setting, our approach addresses the joint sub-band assignment and power allocation problem in cellular vehicle-to-everything (V2X) communications. We propose a novel two-stage systematic explainability framework leveraging feature relevance-oriented XAI to simplify the DRL agents. While the former stage generates a state feature importance ranking of the trained models using Shapley additive explanations (SHAP)-based importance scores, the latter stage exploits these importance-based rankings to simplify the state space of the agents by removing the least important features from the model input. Simulation results demonstrate that the XAI-assisted methodology achieves 97% of the original MADRL sum-rate performance while reducing optimal state features by 28%, average training time by 11%, and trainable weight parameters by 46% in a network with eight vehicular pairs.

Figures (9)

• Figure 1: System model for the V2X communication network.
• Figure 2: Illustration of the proposed multi-agent deep reinforcement learning algorithm.
• Figure 3: The process of generating SHAP_values for multiple trained models using Deep-SHAP. The pre-trained models $f_k(\mathbf{x}), \forall k \in K$ and the background dataset $\mathcal{X}_{BG}$ are utilized to calculate SHAP_values for each instance in the hold-out dataset $\mathcal{X}$. The $\textit{argmax}$ output helps select the appropriate vector of SHAP_values. Subsequently, the mean-absolute SHAP_values are calculated and sorted in descending order to encode feature importance. These values are then transformed using the softmax transformation. The transformed SHAP_values are averaged across the $K$ trained models. Here, $L$ is the total number of state features, and $M$ is the number of valid actions.
• Figure 4: Average network performance and retained state features versus the precision threshold, with: (a) $K=N=4$ where the Original-MADRL network has 24 state features per agent, and (b) $K=N=8$ where the Original-MADRL network has 80 state features per agent.
• Figure 5: Training and testing performance of network with $K = N = 4$, $\boldsymbol{\varepsilon}_{\max}=[\varepsilon_{\text{max}, 1}, \dots,\varepsilon_{\text{max}, k}, \dots, \varepsilon_{\text{max}, K}]=$$[10^{-4}, 10^{-3}, 10^{-4}, 10^{-3}]$ , where (a) shows the learning process comparison for the different algorithms, (b) shows the empirical CDF of network sum-rate performance. Each value is the moving average of the previous 300 time slots.