[DRL on Airsim environment : DO WE NEED Test environment also?]

[Charles-Lim93]

Dear Airsim masters

I'm doing Deep Reinforcement Learning on the Airsim environment, here's the issue.

I ran sample RL in Airsim, (https://github.com/Microsoft/AirSim/tree/main/PythonClient/reinforcement_learning).

And thankfully It worked well for training, but the original code do not have an evaluation code.

So I put
model = DQN.load("dqn_airsim_policy", env = env) mean_reward, std_reward = evaluate_policy(model, model.get_env(), n_eval_episodes=5) print(mean_reward, std_reward)

for example, there are 7 action spaces in a discrete environment, and there was no problem during training.
But when evaluating policy, the action comes out like episode 1's action => 1/2/2 episode 2's action => 1/2/2.... episode 3's action> 1/2/2..... and so on.

Is there anyone who has the similiar issue? Can anyone share their experience on DRL on Airsim?

Thank you in advance.

My environments

• Stable-baseline3
• Airsim 1.6
• Python 3.7
• GTX 2080 Ti

[Charles-Lim93]

Update:

I found out that it is due to lack of not enough training set. Thanks all;

[Gavin2417]

hi, did you manage to get env = VecTransposeImage(env) working, which is a part of the dqn_car.py?