You can account for a good deal of the sub-optimal through artificially generating levels of thinking. Player A’s AI can look at player B’s discard pile, revealed cards and hand size to figure out what possible options and the optimal play weight for player B. Player A then weighs their decision based on both player A’s optimal move and player B’s most probable options, without actually cheating and looking at player B’s hand. Likewise, player B factors player A’s optimals into defining their own range, so both AI are able to skew their combat reveal based on the opponent’s most dangerous options. This should in theory mean that the AI players will add sub-optimal plays into their range or increase their weight from turn to turn based on the game state, over and above whatever data was fed into the DeepMind calculations. I wouldn’t say I’m an AI expert or anything, just throwing out the suggestion.