Customized Action Selection in RL DQN

Hi,

I would like to ask if the latest Reinforcement Learning (RL) toolbox version supports customized action selection.

I’m currently using a DQN agent, and the action in each time step is selected randomly following the epsilon-greedy algorithm. However, I would like to feed in some probabilities in the action selection, such that certain actions are more likely to be chosen. Is this possible using the RL toolbox?

Thank you!

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

I believe this is not possible yet. A potential workaround (although not state dependent) would be to emulate a pdf by providing actions with higher probabilities multiple times when creating your action space with rlFinitesetSpec but I haven't tested that. So something like:

actInfo = rlFiniteSetSpec([-2 0 2 2 2])

[ches]

Hello,

Thank you for the information.

I'm currently trying to improve the exploration during training, so I'm thinking of other ways to do that apart from adjusting the epsilon parameters of the epsilon-greedy algorithm.

In line with that, may I also ask if the following are possible in the latest RL toolbox?

- Setting optimistic initial values

- Other exploration strategies (such as Boltzmann)

Thanks!