Open slerman12 opened 5 years ago
@slerman12 Show the output of your training log, the command you ran, and how long you trained for. Hard to say without seeing these details.
I used this tutorial code based on Baselines: https://github.com/simoninithomas/Deep_reinforcement_learning_Course/tree/master/A2C%20with%20Sonic%20the%20Hedgehog
agent.py runs the code. I just changed the sonic environments to 8 copies of Pong. His learn method in model.py also divided up the training update into mini batches since his Runner was collecting a lot more data for Sonic than could be processed all at once; I changed that back to 1 minibatch since these 8 Pong environments don't demand quite so much data. I also brought his nsteps down to different values, including 5 which is the default in the Baselines code here. No matter what, the model did not seem able to learn to play Pong. The explained variance went up, but the average reward returned per training episode continued to be negative after two days of training.
Hello, If you are looking for training script + working hyperparameters, I recommend you taking a look at the rl zoo (there are also pre-trained agent there).
Disclaimer: this is using stable-baselines, a fork of OpenAI Baselines, but the underlying implementation is the same.
I'm trying to run A2C, but the default parameters, including nsteps=5, aren't letting me train Pong. What parameters to the model can I pass that would let me play a simple Atari game like Pong?