value -0.0000; Nash strategy of the row player [0.5  0.25 0.25]; exploitability at Nash 2.22e-16

regret matching
       T   expl(average)   expl(last)  max avg regret    bound
      10         0.38802       4.0000         0.19401   2.1909
     100         0.16527       2.0000         0.08263   0.6928
    1000         0.04538       2.9241         0.02269   0.2191
   10000         0.01573       2.0000         0.00787   0.0693
  100000         0.00500       2.0000         0.00250   0.0219

fictitious play (exploitability of the empirical averages)
      10         0.61538
     100         0.13592
    1000         0.05184
   10000         0.01600
  100000         0.00464

Hedge with a fixed step 0.1
       T   expl(average)   expl(last)
      10         0.44402       0.3600
     100         0.05615       0.4944
    1000         0.03257       1.5055
   10000         0.01398       1.9994
  100000         0.00446       4.0000

softmax gradient play against a uniformly drawn past opponent iterate (population), step 0.1
       T   expl(average)   expl(last)
      10         0.46346       0.3818
     100         0.40861       1.8739
    1000         1.94354       2.0000
   10000         0.91498       2.0000
  100000         1.02930       4.0000

Shapley's non-zero-sum game, fictitious play: row player's empirical frequencies over t in (1e5, 2e5], sampled every 1000
  min of each coordinate: [0.136 0.152 0.292]  max: [0.269 0.572 0.578] (a Nash equilibrium has all coordinates 1/3)
