I'm testing various learning rates and neural network configurations. I'm testing over 10000 games, with the first 2000 having random starting moves and general randomness throughout of about 20%, i.e. 20% of moves are random.

In all configurations, I initiate the weights to random values.

What I've found is that in all configurations, Player 1 will win the majority of games, or Player 2 will. There's no 50/50 split.

Is this expected or normal? I am training on the Connect Four game.

  • $\begingroup$ This is weird, how big are the majorities? I would also decrease your epsilon (probability of random move) as training progresses. Do you test with an epsilon? $\endgroup$ – Jaden Travnik Dec 14 '17 at 3:11
  • $\begingroup$ @jaden What are "the majorities", I'm not familiar with this term. I do reduce the randomness over time, it starts at something like 50% and drops to 10% after the first 10% of moves. 20% was way off. $\endgroup$ – NeomerArcana Dec 14 '17 at 4:13
  • $\begingroup$ By majorities i mean by how much does player 1 beat player 2 (eg 90%) and vice versa. $\endgroup$ – Jaden Travnik Dec 14 '17 at 12:55
  • $\begingroup$ It's hard to tell, it's a win or lose, there's no score or anything. $\endgroup$ – NeomerArcana Dec 14 '17 at 19:07
  • $\begingroup$ Sorry, I meant in regards to what you said about there is no 50/50 split. What is the split? Like 90/10? $\endgroup$ – Jaden Travnik Dec 14 '17 at 23:53

From wikipedia article on Connect Four:

Connect Four is a solved game. The first player can always win by playing the right moves.

It's pretty reasonable that even without training 1st player win more often by small margin. With some training 1st player should mostly win.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.