Year: 2025
Type: article
Abstract: Deep Reinforcement Learning (DRL) has achieved remarkable success in optimizing complex control tasks; however, its opaque decision-making process limits accountability and erodes user trust in safety-critical domains such as autonomous driving and clinical decision support. To address this transparency gap, this study proposes a hybrid DRL framework that embeds explainability directly into the learning process rather than relying on post hoc interpretation. The model integrates symbolic reasoning, multi-head self-attention, and Layer-wise Relevance Propagation (LRP) to generate real-time, human-interpretable explanations while maintaining high control performance. Evaluated over 20,000 simulated episodes, the hybrid framework achieved a 91.9% task-completion rate, a 19.1% increase in user trust, and a 15.3% reduction in critical errors relative to baseline models. Human–AI interaction experiments with 120 participants demonstrated a 25.6% improvement in comprehension, a 22.7% faster response time, and a 17.4% lower cognitive load compared with non-explainable DRL systems. Despite a modest ≈4% performance trade-off, the integration of explainability as an intrinsic design principle significantly enhances accountability, transparency, and operational reliability. Overall, the findings confirm that embedding explainability within DRL enables real-time transparency without compromising performance, advancing the development of scalable, trustworthy AI architectures for high-stakes applications.
Source: Mathematics
DOI: 10.3390/math13233805
Authors: Wassim Jaziri, Najla Sassi
Institution: King Faisal University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
Funder: King Faisal University
OpenAlex ID: W4416746784