btaba
commented
7 years ago Hi lvlvlvlvlv, to asses performance of the training process you can take the average return per training episode, and plot that over time. Although with the MC method you are not ensured that the returns monotonically increase, you should generally see the average return increasing by training episode. Alternatively you can make an agent which performs random actions, and plot the average return of that agent over time. Then you can compare the random-agent plots with the MC-agent to assess performance. Hope that answers your question!
xubo92
@btaba hi bro: I am on the task of racetrack problem in Sutton's RL book with the method of on-policy monte carlo. However, I find it hard to assess the performance of training process. Is there any parameter which can reflect the performance of training process? I also tried to make a movie but found that the training process is hard to assess in that way. I cannot ensure the relationship between episode number and training performance. For instance. if your performance looks bad, how to tell that the key is lack of episode number or just your wrong implementation of algorithm? Very confused about those questions. Do you have any ideas? Look forward to your reply!