COIN Framework

The goal is to find local utility functions for the agents such that the global utility is maximized, assuming the agents are using reinforcement learning.
It has three main components:
1. Defines equivalence classes of local utility functions such that if they increase so does the global utility.
2. Further restricts the allowable utility functions such that agents do not frustrate each other by lowering others' utility. The design for these first two is the COIN initialization.
3. Modify the local utility functions at runtime based on localized statistical information in order to better align them to the global.
Microlearning refers to each agent's reinforcement learning.
Macrolearning refers to the modification of each agent's utility function (step 3).