question about calculate loss value of dqn agent reinforcement learning

3 次查看（过去 30 天）

Kun Cheng 2023-6-25

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1987883-question-about-calculate-loss-value-of-dqn-agent-reinforcement-learning

编辑： Emmanouil Tzorakoleftherakis 2023-6-27

hello everyone,

I want to plot curve of loss value during training. It is not available in dqn training option directly as i know. My question is :

can i calculate loss simply with this function?

Q0 is predicted value by critic. (orange line in default process-training plot)

EpisodeReward is gained reward by agent in t episode

t is number of episode

%% example code for loss plotting here
% unpack training statistic
index = trainingStats.EpisodeIndex;
reward = trainingStats.EpisodeReward;
pred_reward = trainingStats.EpisodeQ0;
Loss_Plot = [];
for t = 1:size(index, 1)
    loss = (sum(pred_reward(1:t, 1)) - sum(reward(1:t, 1)))/t
    Loss = [t, loss];
    Loss_Plot = [Loss_Plot; Loss]
end
% plot curve
plot(Loss_Plot(:, 1), Loss_Plot(:, 2), '-')