Dataset and Trained Models: Social Information Quality and Environmental Volatility Shape Collective Foraging Behavior

Abstract This dataset contains the trained multi-agent reinforcement learning (MARL) policies, training parameter logs, and aggregated evaluation rollouts for the study "Social Information Quality and Environmental Volatility Shape Collective Foraging Behavior." Because the neural network weights and evaluation rollouts are substantial in size, they are hosted here separately from the main codebase. This data allows researchers to bypass the computationally expensive training phase (which required up to 240 million environment interactions per run) and directly reproduce the behavioral analysis, figures, and videos presented in the manuscript. Dataset Contents wandb-logs.zip: Contains the raw artifacts generated during the MARL training phase in multiple folders. This includes the PyTorch model weights (.pth files) for the actor-critic networks at various training iterations (e.g., step 400 to 480) and the specific environmental hyperparameters (run_params.json) for every experimental condition and random seed. Naming Convention: Folders within this archive are named using run hyperparameter. Example folder: target_0.5-siq_distance-vis_range_15-training_seed_0-tracking_time_cost_0.9-shared_parameters_True Explanation: target_0.5 = Resource speed (environmental volatility). siq_distance = Social Information Quality condition (e.g., distance, payoff_no_noise). vis_range_15 = Visual range of the agents. training_seed_0 = Random seed used for the run. tracking_time_cost_0.9 = Cost of private tracking (effective speed multiplier). shared_parameters_True = Indicates centralized training (True) vs. decentralized (False). evaluation.zip: Contains the output generated by the policy evaluation scripts (evaluation.py and aggregate_evaluation_results.py). Naming Convention: Folders within this archive are named using run hyperparameter. Example: 0.9-True-0.5-siq_distance-15-480-0 Explanation: 0.9 = Tracking time cost. True = Shared parameters (centralized learning). 0.5 = Target/Resource speed. siq_distance = Social Information Quality condition. 15 = Vision range. 480 = The evaluation step/iteration. 0 = Training seed.