Optimal Accelerated Life Testing Design Under Constrained Resources Using Double Deep Q‐Learning

ABSTRACT Accelerated life tests (ALTs) are essential tools for estimating product reliability under high‐stress conditions, allowing failure data to be collected in reduced timeframes. However, planning effective ALT configurations is a complex task that requires selecting stress levels, test durations, and unit allocations while accounting for limited resources and model uncertainties. This study proposes a reinforcement learning (RL) approach to automate and optimize test planning under a failure model that emulates real‐world reliability patterns. The test environment simulates product failures using a mixture of Weibull distributions modulated by an Arrhenius temperature scaling, yielding possibly different hazard functions, such as the power‐law increasing and decreasing curves, constant, and the bathtub‐shaped curve, typical in industrial applications. A Double Deep Q‐Network (DDQN) agent is trained to sequentially configure test levels by selecting temperature, time, and sample size, with the goal of minimizing the statistical uncertainty in estimating the thermal sensitivity parameter , which is crucial for extrapolating lifespan predictions under normal operating conditions. The trained agent consistently produced shorter test plans that enabled precise estimation of all model parameters, notably achieving lower variance in the estimate compared to competing methods. These results suggest that the proposed RL framework offers a flexible and adaptive alternative to traditional experimental designs.