We have recently improved the design of saving agents with relevant training information. In the new design (available from R2022a), the saved agent MAT file will contain the agent (under the variable "saved_agent") and a training result object (under the variable "savedAgentResult"). The training result object replaces the previous "savedAgentResultStruct". We will fix the example to reflect the change.
The example should still work fine when comparing the existing "DDPGAgent" and "TD3Agent" folders. For new trainings, you can replace lines 76-82 in comparePerformance.m with the following code:
for i = 1:size(files,1)
S = load(files(i).name);
if isfield(S,'savedAgentResultStruct')
res = S.savedAgentResultStruct.TrainingStats;
info = S.savedAgentResultStruct.Information;
else
res = S.savedAgentResult;
info = res.Information;
end
agent.averageReward(:,i) = res.AverageReward;
agent.EpisodeQ0(:,i) = res.EpisodeQ0;
agent.EpisodeIndex = res.EpisodeIndex;
agent.ElapsedTime = info.ElapsedTime;
end
You can retrieve the elapsed time from the training result object using
t = savedAgentResult.Information.ElapsedTime
Do remember that for each new training you will need to extract the MAT-file of interest and place in a separate folder other than the auto-generated "savedAgents" folder before doing the comparison.
