DDQN isn't learning as expected

[wxue24]

The reward plot as shown above is decreasing rather than increasing over time. Could be due to hyperparameters chosen, or how the state features are preprocessed.

Some ideas to try:

• Preprocess data (normalize values like position to between 0 and 1)
• Experiment with different neural network layers
• Experiment with different hyper parameters like "burnin", "learn_every", "sync_every", and "save_every", "batch_size", "exploration_rate"

[wxue24]

After adding preprocessing to the data and adjusting the hyperparameters I got a much better reward plot.

We can try to make more adjustments to decrease the training time but it seems like the result is pretty good as the end reward is > 100.