Reinforcement Learning Tool Box : How to change epsilon during training?

I just started using reinforcement learning tool box on matlab.

I have a question in terms of epsilon-greedy settings. Is it possible to change the value of epsilon during training?

For instance, for the firtst 30 episodes, I'd like to set Epsilon = 1 and EpsilonDecay = 0.1. Then, after the first 30 episodes, I'd like to change Epsilon = 0.5 and EpsilonDecay =0.05.

I'd really appreciate your comments and advices. Thanks

Answer 1

[Emmanouil Tzorakoleftherakis]

Hi Keita,

Have a look at this link. The 'EpsilonGreedyExploration' option provides a way to reduce exploration as training progresses using the formula Epsilon = Epsilon*(1-EpsilonDecay).

We are working on adding more ways to set up the exploration strategy.