Model training issues

[aLadNamedPat]

Reward is not increasing as the model trains which indicates that there is something wrong with the training process. Possible errors include:

• Tracked grid input is not well defined enough to aid model in finding the correct areas to travel to.
• Algorithm is implemented correctly so learning is difficult (?)
• Steps are not run for long enough per episode for model to learn the environment that it's actually exploring
• Something wrong with the hyper-parameters alpha, beta, tau (least likely reason)

[aLadNamedPat]

The algorithm presented also does not necessarily guarantee that the average number of detections per second. Instead, the final output of episodic reward is the last reward multiplied by step (n+1) / timesteps which means that there is high variance in what the average detections per second is.

This estimate is almost arbitrary so to measure model performance, reward should not be used as the actual metric.