Arcade-style rewards system

[bzier]

Idea for an alternate rewards system:

Old school arcade game style checkpoint system

• Every step is still -1 reward

  any positive reward for steps will likely encourage delays rather than shortest/fastest path - e.g. driving around aimlessly to accumulate reward

• Agent only has x steps to play (some reasonably small number)
• Hitting a checkpoint extends play for another y steps
• Checkpoints also grant a 'large' sum of reward points

  note that the checkpoint reward needs to be enough to make progress worthwhile; if an episode is lengthened, more steps result in a lower total reward; the checkpoints must offset this or the agent may learn to maximize reward by simply avoiding the first checkpoint and never extending an episode

In theory, this system should/could:

• reduce overall episode lengths (by terminating early if no progress is being made), which allows for shorter/faster iterations (i.e. fail fast)
• reduce getting stuck in any one place for too long
• prevent driving backwards too far
• still encourage forward progress with checkpoints as before

[bzier]

Depends on #26. I would like to see the reward functions injectable/pluggable to facilitate experimentation with all sorts of variations. Before working on this implementation, there should be a way to easy swap out which reward function should be used.