Early NashConv Notes¶
Historical scratch note. See NashConv Reporting for the implemented estimator, report schema, and confidence semantics.
Approximate Unilateral Deviation Incentive¶
Pick your evaluated joint policy \(\pi\). For each player (i):
- Freeze opponents \(\pi_{-i}\)
- Train a “deviator/exploiter” policy \(\hat{\pi}*i\) in the environment to maximize (i)’s return against \(\pi*{-i}\)
- Estimate by rollouts:
- \(V_i(\pi)\): everyone uses \(\pi\)
- \(V_i(\hat{\pi}*i, \pi*{-i})\): player (i) deviates
Approximate NashConv¶
Same as the above?