Open jlabhishek opened 5 years ago
First of all, if your deployment environment is vastly different from what we trained, you will have to retrain the agent in your new environment. Also, there is actually a deeper problem for training with a large number of videos with different length. The variance in the video length (hence the MDP horizon and scale of total reward) will introduce extra variance that might hurt the RL algorithm. We recently show this effect has broader impact in a wide range of environments: https://arxiv.org/abs/1807.02264
Hi, Can you share some insight regarding what kind of results we can expect when comparing the results for different algorithms for a video which has different chunk duration and total chunk count.
I tried and the results were not very good for me, I changed the values in rl_no_training.py and fixed_env.py , could have I missed something?
Thanks