Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

GAE and Critic Loss (PPO2) #1190

Open
rohey opened this issue Oct 11, 2021 · 0 comments
Open

GAE and Critic Loss (PPO2) #1190

rohey opened this issue Oct 11, 2021 · 0 comments

Comments

@rohey
Copy link

rohey commented Oct 11, 2021

Hello. Can you please explain why are you using the mb_returns = mb_advs(GAE) + mb_values as the returns to compute the critic loss ? Should not the value function approximately represent the discounted sum of rewards ? E.g., R = gamma * R + rewards[i] value_loss = value_loss + 0.5 * (R - values[i]).pow(2).

If I understand correctly, the value function depends on the parameter γ and not on the parameter λ based on the paper https://arxiv.org/pdf/1506.02438.pdf. However, If I use the GAE(γ, λ) advantages to compute the returns and use those returns to train the critic, wouldn't the the value function become V(γ, λ) instead of V(γ) ? And if it is true, would the TD residual be correctly computed with V(γ, λ) ?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant