eleurent
commented
1 year ago Hey! I pushed a fix, you can try again and let me know if this solves your issue.
In addition, I would like to ask if you have achieved 100% safety purely relying on reinforcement learning. I found in the highway documentation that you use DQN from stable_baselines3 to make decisions, but there are no related performance indicators such as collision rate and average speed. So I would like to ask if it is feasible to rely solely on reinforcement learning to achieve automatic driving decision-making.
In my own experience, yes it is feasible to achieve "100% safety" (as evidenced by an empirical 0 collision out of some number of episodes, not a proof per se) especially in simplified simulated domains like highway-env. The real world is another beast. And of course, it also depends on a number of factors and design choices: what to use for observations, for actions, the reward function (usually a safety vs efficiency tradeoff is involved here), the network architecture, the high level algorithm, the implementation, etc. I refer to stable_baselines3 in the docs because it is a strictly better library than this one (better maintained, lets of algorithms, wide community), but I mostly experimented with rl-agents. You can look at my PhD publications for more details. Unfortunately I focused mostly on reward maximisation which aggregates safety metrics like collisions vs efficiency metrics like speed, and I did not always provide a table distinguishing the two, which I agree would have been a great idea.
zh1114haoshen
Hello Mr. Edouard, first of all, thank you very much for your contribution to the automatic driving in the highway scenario. I am very interested in this part. My current job is to deploy reinforcement learning to the highway scenario, using the intersection_social_dqn.ipynb framework in the highway repository. I just changed the environment to the highway environment and deployed DQN, DDQN, and duelingDDQN without any problems.
However, when I deployed FTQ to the highway scenario, I encountered an attribute missing problem, which prompted： AttributeError: 'Evaluation' object has no attribute 'seed'.
I don't know how to solve it and I am looking forward to your reply.
In addition, I would like to ask if you have achieved 100% safety purely relying on reinforcement learning. I found in the highway documentation that you use DQN from stable_baselines3 to make decisions, but there are no related performance indicators such as collision rate and average speed.
So I would like to ask if it is feasible to rely solely on reinforcement learning to achieve automatic driving decision-making.