This paper develops a generalizable RL approach in which the agent learns effective policies to achieve risk-informed decision-making (RIDM) objectives. While the RL approach can be applicable to various RIDM contexts, this paper specifically focuses on reducing system risk below a predefined acceptance criterion with the minimum number of design modifications, which can arise in risk-informed design of nuclear reactors.

The RL environment is implemented using the Gymnasium package in Python and is defined by a risk model expressed as the union of minimal cut sets, paired with a discrete action space. Available actions include identifying critical components using importance measures and improving their reliability. The agent is trained using Q-learning, which iteratively refines action selection to minimize the risk function in as few modification steps as possible.

Results from multiple replications demonstrate consistent convergence of the RL agent training: the trained agent reaches the target risk threshold in the same number of steps and with similar action sequences across runs. This observation confirms that Q-learning identifies a stable, reproducible policy. This paper extracts preliminary generalizable insights from the trained agent's behavior, including which importance measures are most informative in the context of system risk reduction and how qualitative structural characteristics of the risk model, such as redundancy level, can supplement importance measure results to guide decision-making. These findings can support the broader goal of developing generalizable RL agents trained across diverse PRA models, such that derived policies can inform RIDM for a range of nuclear energy systems rather than a single, system-specific configuration.