Training Process Dataset for the Radioactive Source Search Method Based on Angle-of-Arrival Particle Filtering and Reinforcement Learning

The training environment is an indoor radioactive source search scenario of 20 m × 20 m, with 1–7 randomly placed rectangular obstacles, source strengths ranging from 15000 to 20000 cps, and background intensities ranging from 50 to 500 cps. The dataset is provided as a progress log (progress.txt) recording 8 key metrics per epoch: Average Episode Return (AverageEpRet), Standard Deviation of Episode Returns (StdEpRet), Number of Successful Episodes (DoneCount), Average Episode Length (EpLen), Policy Entropy (Entropy), KL Divergence (KL), Model Loss (LossModel), and Value Function Loss (LossV), totaling 3000 rows, with each row corresponding to one training epoch.

During training, the success rate of the agent increased from approximately 2.5% at initialization to about 97.8%, and the average episode return converged from -116 to +16.5, approaching the theoretical maximum (success reward = +20). The entropy, KL divergence, and all loss metrics converged normally, which supports the training-effectiveness analysis in the paper.