As it is the code can't run because of little problems like use of the Nim class and the population list before their initialization, use of a "limit" variable which is never defined. With these small changes everything works fine.
Task 2.1 hasn't been completed as there's no expert strategy different from the one proposed as optimal.
Task 2.2: (mu+lambda)-ES doesn't usually use recombination and mutation probability, while parent selection should be completely random.
Implementation of the k bound variant of the game is missing.
The selection of "good offsprings" when the best individual in them is better than a certain limit-th individual in the previous population isn't an orthodox way to choose which individual are going to be kept every generation.
I tested it and often a new best individual would perform way worse in the next generation, probably an higher value of the "accuracy" variable could make each individual's performance a bit more stable between tests.
A good addition could have been to alternate the order of players in the tournament when computing individual's fitness (Nim favors the one who plays first or second based on the dimension of the match and consequential initial nim-sum).
It would also be nice to see some data about the results, maybe with different paramenters (like "accuracy" and "population_size") to show different performances.
As it is the code can't run because of little problems like use of the Nim class and the population list before their initialization, use of a "limit" variable which is never defined. With these small changes everything works fine.
Task 2.1 hasn't been completed as there's no expert strategy different from the one proposed as optimal.
Task 2.2: (mu+lambda)-ES doesn't usually use recombination and mutation probability, while parent selection should be completely random. Implementation of the k bound variant of the game is missing. The selection of "good offsprings" when the best individual in them is better than a certain limit-th individual in the previous population isn't an orthodox way to choose which individual are going to be kept every generation. I tested it and often a new best individual would perform way worse in the next generation, probably an higher value of the "accuracy" variable could make each individual's performance a bit more stable between tests. A good addition could have been to alternate the order of players in the tournament when computing individual's fitness (Nim favors the one who plays first or second based on the dimension of the match and consequential initial nim-sum). It would also be nice to see some data about the results, maybe with different paramenters (like "accuracy" and "population_size") to show different performances.