When the neurocontroller is embedded within a game environment, its training is complete and no further learning is possible. To include the ability for the character to learn, portions of backpropagation could be included with the game to adjust weights based upon game play.
A very simple mechanism could adjust the weights of the neurocontroller based upon the last action made by the character. If the action led to negative consequences, such as the death of the character, the weights for this action given the current environment could be inhibited to make it less likely to occur in the future.
Further, all game AI characters could learn these same lessons, in a form of Lamarckian evolution (whereby children inherit the traits of their parents, which would include lessons learned). After numerous games played , the game AI characters would become slowly better at avoiding negative situations.