IMP-MARL dataset

IMP-MARL dataset Authors: Pascal Leroy ([e-mail]), Pablo G. Morato ([e-mail]), Jonathan Pisane, Athanasios Kolios, Damien Ernst GitHub repository: https://github.com/moratodpg/imp_marl IMP-MARL environments:<br> struct_uc: k-out-of-n system<br> 2-out-of-3, 4-out-of-5, 9-out-of-10, 48-out-of-50, 95-out-of-100<br> (with and without campaign cost)<br> struct_c: correlated k-out-of-n system<br> 2-out-of-3, 4-out-of-5, 9-out-of-10, 48-out-of-50, 95-out-of-100<br> (with and without campaign cost)<br> owf: offshore wind farm<br> 1 wind turbine, 2 wind turbines, 5 wind turbines, 25 wind turbines, 50 wind turbines<br> (with and without campaign cost) MARL methods investigated: QMIX, QVMIX, QPLEX, COMA, FACMAC, IQL, DQN. This dataset contains:<br> MARL logs.zip =&gt; The configuration, execution, and result files of all investigated MARL methods.<br> Each zip file contains the logs from an IMP-MARL environment.<br> There, test and train files are stored for each method:<br> cout.txt contains stored console ouputs.<br> config.json contains the list of all parameter values for the experiments.<br> info.json is the log file, listing reward values, for example.<br> run.json contains simply logged information about the host, typically hardware.<br> best_agent_networks.zip =&gt;The controller networks’ weights of the best MARL-based policies. <br> Controller files are stored as a pytorch file for each method/environment investigated (agent.th).<br> heur_logs.zip =&gt; The configuration, execution, and result files of the computed expert-based heuristic policies. <br> /heur_search =&gt; logs resulting from all evaluated heuristic policies, stored as numpy files.<br> /heur_search =&gt; results from the best performing heuristics over 10,000 policy realisations, stored as pickle files.