Although the eTemp wasn’t being reset in the agent_init method, it wouldn’t have mattered in this case. In the proving run though, it would have been quite bad.
Anyway, the results obtained appear much the same, so clearly the rate of learning is the problem.
There are three different areas to change: The Beta value (multiplier of mutations), the Cooling Rate, or the initial parameter, which may be overfit to the first MDP.
Currently testing 2 other values of Beta to see if the results are any better/worse.
To test the other parameters, I could use the best Beta from these tests and trial different cooling rates. For the initial parameter, to see which value is ideal, I could run a 0.5 eTemp agent to see what parameters are best overnight (using {1,1,1,1,1} as the initial parameter).
Presentation is coming up…