Merge pull request google-deepmind#1049 from rezunli96:nash_avg

PiperOrigin-RevId: 524798649 Change-Id: I7c791d53abc056865448e83edff0058f0ae9dccc
sandguine · Apr 17, 2023 · 1e0c18e · 1e0c18e
2 parents ca8affc + 640bad8
commit 1e0c18e
Showing 1 changed file with 6 additions and 3 deletions.
diff --git a/open_spiel/python/algorithms/nash_averaging.py b/open_spiel/python/algorithms/nash_averaging.py
@@ -82,8 +82,11 @@ def nash_averaging(game, eps=0.0, a_v_a=True):
   # game does not have to be symmetric
 
   m, n = p_mat[0].shape
-  a_mat = np.block([[np.zeros(shape=(m, m)), p_mat[0]],
-                    [-p_mat[0].T, np.zeros(shape=(n, n))]])
+  min_payoffs = np.min(p_mat[0], axis=1).reshape((m, 1))
+  max_payoffs = np.max(p_mat[0], axis=1).reshape((m, 1))
+  std_p_mat = (p_mat[0] - min_payoffs)/(max_payoffs-min_payoffs)
+  a_mat = np.block([[np.zeros(shape=(m, m)), std_p_mat],
+                    [-std_p_mat.T, np.zeros(shape=(n, n))]])
   maxent_nash = np.array(_max_entropy_symmetric_nash(a_mat, eps=eps))
   pa, pe = maxent_nash[:m], maxent_nash[m:]
-  return (pa, pe), (p_mat[0].dot(pe), -p_mat[0].T.dot(pa))
+  return (pa, pe), (std_p_mat.dot(pe), -std_p_mat.T.dot(pa))