Neural_Evolver, after 18 hours of training, seems to fixate on choosing pure strategies. That is, it chooses 14 (at least 90% of the time) given initial conditions. Perhaps this became the norm because choosing 14 was better than the uniform-like player. But it's surprising that Neural_Evolver jumped to always choosing 14 in this case. (Did it jump? Or is there a bug?) If there's no bug, a solution seems to be to have a larger population (like 15) and have real, random evolution so that 14 becomes common and then less so because it must face itself and the uniform-like player. But maybe the neural network I'm using is unsuited for evolution. Maybe its output is not continuous based on the parameters, or something. This issue can wait until neural_nash is operating.
Neural_Evolver, after 18 hours of training, seems to fixate on choosing pure strategies. That is, it chooses 14 (at least 90% of the time) given initial conditions. Perhaps this became the norm because choosing 14 was better than the uniform-like player. But it's surprising that Neural_Evolver jumped to always choosing 14 in this case. (Did it jump? Or is there a bug?) If there's no bug, a solution seems to be to have a larger population (like 15) and have real, random evolution so that 14 becomes common and then less so because it must face itself and the uniform-like player. But maybe the neural network I'm using is unsuited for evolution. Maybe its output is not continuous based on the parameters, or something. This issue can wait until neural_nash is operating.