Reward Tuning about rect-agent-scripts

[choyai]

New Reward Scheme

score -> 1 - out of bounds reward( the further out, the lower reward
lose -> -0.5 - out of bounds reward, clamped to -1 as a minimum value, and 0 as a maximum value

Hopefully this will have an effect of incentivizing agents to stay inside boundaries.

[choyai]

this issue is still not resolved completely it seems. I am observing inconsistent behavior between teams.

[choyai]

visit this link to see the tensorboard results for this week's training.
The run sense did better than the other two runs, because I fixed a bug where the observations were relative to world instead of agent. As referenced above, though, the blue agents still behave oddly.

[choyai]

These videos show how the agents performed.

run-id: skill

Blue always moved away from the field, so I added a penalty to going outside the boundaries.

run-id: tune

After this run, I realized something was wrong with the sensor inputs as blue is still always moving away from the net. I changed so that all observations are relative to the agent.

run-id: sense

There's still a positioning problem, albeit opposite the other runs.

[choyai]

run-id: aim

no clip, but ELO dropped extremely quickly. Agents tended to avoid the ball after service.

run-id: hitreward

Added a small reward for hitting the ball, which meant agents were incentivized to hit the ball, but it turned out to only improve the ELO by a bit. The agents quickly settled into the same behaviour, avoiding the inside of the area.

[choyai]

After a while, the agents tended towards the same policy: running away.

[choyai]

What's Next?

There are currently 2 paths I've identified for training.

Straight Shot( No fixed skill ) path of training

I've found the bug that forced the agents to learn bad behaviour: the environment would reset itself at seemingly random after service. Each agent earns no rewards during the process, and thus, can learn bad behaviour.

If we continued the straight shot path of training, I would just upgrade the ml-agents version and refactor the codebase a little bit, then start training right away.

Fixed Skill

Unfortunately, I wasn't able to find examples of MARL trained agents with fixed skill, other than in ai-economist

This approach requires an implementation of how skill works, which, given that ai-economist just specifies it as 'skill', I would need to model on my own. The leading candidate solution right now is the concept of a skill graph, with attributes defining the skill an agent has.

For volleyball, I've specified 3 attributes to contribute to skill:

1. Power
2. Accuracy
3. Speed

These can easily be implemented for the agent to use an hopefully it will help speed up training.