Skip to content

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):

  1. Freeze opponents \(\pi_{-i}\)
  2. Train a “deviator/exploiter” policy \(\hat{\pi}*i\) in the environment to maximize (i)’s return against \(\pi*{-i}\)
  3. Estimate by rollouts:
  4. \(V_i(\pi)\): everyone uses \(\pi\)
  5. \(V_i(\hat{\pi}*i, \pi*{-i})\): player (i) deviates

Approximate NashConv

Same as the above?