[Python] New to DQN, trying to train a Lunar Lander(v3) model, but my rewards are not...

I am very new to reinforcement learning and trying to train a model for Lunar Lander for a guided project that I am working on. From the training graph (reward vs episode), I can observe that there really is no improvement in the performance of my model. It kind of gets stuck in a weird local minima from where it is unable to come out. The plot looks like this:



I have written a jupyter notebook based on the code provided by the project, where I am changing the environments.

The link to the notebook is this.

I am unable to understand what is (if there is anything wrong with this behavior, and if it is due to a bug in the code). Because I feel like, for a relatively starter environment, the performance should be much better and should increase with time, but it does not happen here. (I have tried multiple different parameters, changed the model architecture, played around with LR, EPS_Decay but nothing seems to make any difference to this behaviour)

Can anyone please help me in understanding what is going wrong and if my code even is correct? That would be a great favor and helped you'd be doing to me.

Thank you so much for your time.

Continue reading...