qasimwani / policy-value-methods

Deep Reinforcement Learning algorithms for Policy Value methods written from scratch.

policy-value-methods

My implementation on bunch of policy value methods from scratch

Algorithms:

1. Hill Climb
2. Cross Entropy Method
3. Policy Gradient Methods
   1. REINFORCE
   2. PPO (Proximal Policy Optimization) Video
   3. Actor Critic

Results:

LunarLander (REINFORCE) {Solved in 519 episodes}

BipedalWalker-v3 (TD3) {completion time ~14seconds, achieved after 500 episodes}

Score

Rolling score

policy-value-methods's Issues

A3C model doesn't converge!

fluctuating losses. agnostic to number of parallel agents. checked loss function, everything seems fine when referenced across A3C paper and other repos. shared optimizer looks fine.
can't figure out the exact issue.