Gain in VPGPolicy does not account for terminal states?

[ArjunNarayanan]

I think we need to pass traj[:terminal] to discount_rewards so that the gain is computed only up to termination of an episode?

ReinforcementLearning.jl/src/ReinforcementLearningZoo/src/algorithms/policy_gradient/vpg.jl

Line 105 in 6fe6aa0

gains = traj[:reward] |> x -> discount_rewards(x, π.γ)

[findmyway]

Agree, nice catch!

How about making a PR?

[ArjunNarayanan]

I'm happy to do it. I've never done it before. Could you guide me through the process? I guess I have to fork the package, make the changes on my local copy in a separate branch?

[findmyway]

Yep, you may find many resources by searching "first github pr" (like this one )