REINFORCE not able to learn policy

Hi everyone,

I´m trying to implement REINFORCE as described in Sutton

However it doesn´t seem to be learning. Bellow is the code for the train function (using Gymnasium CartPole-v1)

´´´
def train(ep, policy, optimizer, gamma=1):
G = 0
for t in range(len(ep)-1):
st, at, rt = ep[t]
G = rt + gamma * G
pi = torch.log(policy(st)[at]) # My NN last layer is a softmax
loss = -G*pi
loss.backward()
optimizer.step()
´´´

What it does is basicaly choose to ways, go for a real bad policy, averaging 10 for the rewards (while a random agente usually scores in the ~20s) or learn a little bit and stay arround 80 avg reward.

First I thought that the policy was changing to abruptly so I inserted the clip_grad_norm_, which increased the amount of time the agent got stucked at the 80ish but didn´t did much else (maybe got stucked in some local minima?)

I tried to play with weight initialization, net archetecture but to no avail. I would appreciate if someone could shed light in what maybe the issue.

Not sure this naive version will do the job if the task is hard-ish… You should perhaps use some more evolved algorithm like PPO or a better advantage estimation like GAE.
One thing I spotted in the code is that you seem to be computing the advange starting from t=0 and ending at t=H, but it should be

``````for t in range(len(traj), -1, -1):
G = r[t] + gamma * G
``````
1 Like

Not sure this naive version will do the job if the task is hard-ish…

Yeah, I was wondering about that, however I was expecting some sort of “asymptotically” learning instead of the plateau that is happening

This Is a “good” run (blue lines the reward for each episode the orange is the moving average)

You should perhaps use some more evolved algorithm like PPO or a better advantage estimation like GAE.

I´m trying to implement the algorithms my self to get a better fealling for how things work, PPO is the next one.