Some thoughts about KAN-based RL

[yuzej]

Thanks god I have seen this repo and found the one working on this idea.
One of my thoughts is that:
the KAN seems like good at the fitting of continuous functions, but DQN, DDQN are discrete action algorithms? Maybe this is the reason of your results. I don't have good sense in fundamental RL, but I think that it maybe useful to find the environments with specialized features which will enlarge the advantages of KAN.
I check the original KAN repo, and the training time is 10x times longer than MLP, and the same result you have made. I'm working on a high-time-resolution dynamic systems and the most thing i really care about is the inference time. Would you please just show us the inference time of your KAN and MLP networks? also the layers details are essential.
Thank you bro!
Looking forward to your reply.

[riiswa]

In fact, we need to explore a lot more and in particular test other environments and other hyperparameters. I'll benchmark the code a bit later, but as an indication, on CartPole I launched a seed in 2 minutes with MLPs, and in 10 minutes with KANs...

[stergiosba]

I read the KAN paper yesterday and did some sketching on a paper on how KANs can be used for interpretable RL. Then I saw that you have already started this repo and I am happy. I think a lot of people are thinking of the same things right now and its really exciting.