Creating a Deep Q-Network which helps the OpenAI gym's classic Acrobot-v1 to learn to cross a line by swinging back and forth on a pole.
The code used is mainly derived from here. A great source for learning about Deep Q-Networks and how to code them. I made a few changes to the code given in the link by using a classic gym environment and also made my own neural network model. Instead of RGB images of the frames of the game, I used direct observations values received from the environment while running each step. These tweaks were done to make the learning computationally less-expensive and to reduce the training time while also getting good results.
After training for 5000 episodes with a small dense neural network, it was able to achieve a 100-episode average result of around -83. (Acrobot-v1 does not have a specified reward threshold at which it's considered solved.)
The number of episodes and the dense neural network architecture can be changed to get better results. Open acrobot_learning_without_img.py for the code.
The model has been saved as main_model.h5. It can be used elsewhere using Keras's model.load_model function and used to play
the games again. The corresponding file doing this is playing_acrobot.py.