Robust AI Decision-Making under Uncertainty using Probabilistic Reinforcement Learning Models

Abstract

The increasing deployment of artificial intelligence (AI) systems in dynamic, complex, and high-stakes environments has highlighted the importance of robust decision-making under uncertainty. Traditional reinforcement learning (RL) approaches often rely on deterministic policies and point-estimate predictions that fail to adequately capture the inherent stochasticity and ambiguity in real-world scenarios. This research introduces a comprehensive framework for Probabilistic Reinforcement Learning (PRL) that explicitly models environmental uncertainty, outcome variability, and policy confidence through probabilistic representations. By integrating Bayesian inference, stochastic value estimation, and uncertainty-aware policy optimization, the proposed approach enhances both the reliability and safety of AI decisions across diverse application domains such as robotics, autonomous driving, healthcare diagnostics, and financial decision support systems.

 The study begins by examining the limitations of conventional RL models that depend on fixed reward functions, stable transition probabilities, and fully observable states—assumptions rarely satisfied in practice. To address these constraints, the research proposes a probabilistic reformulation of key RL components, including transition models, value functions, and policy distributions. Leveraging Bayesian Q-learning, Monte Carlo Dropout-based exploration, and probabilistic policy gradients, the framework captures uncertainty in model predictions and uses it as a signal to guide safer and more informed decision-making. Furthermore, the model integrates distributional reinforcement learning, enabling the estimation of full return distributions rather than single expected values, thereby improving risk-sensitive reasoning and robustness against outliers or rare events.