Conclusion

The math in these papers can be hard to understand because it uses physics notation and tries to be very general.
However, the two main results are easy to understand and apply. When choosing a reward function for a reinforcement-learning MAS, you should:
1. Use the wonderful life utility. That is, give each agent a reward the is the difference of the global reward given the current state and the global reward if the agent had done some other fixed action.
2. Try to make that other fixed action the agent's "average action"
It remains to be seen how (or if) well COIN can be applied to more complex domains.