I have redesigned the AI again to improve its strength further. Here's a quick overview of the new method:

AI players now use a search algorithm to look up to three moves ahead. For each possible sequence of up to three moves, the expected change to the number of dice of that player, including the reinforcements they would get for ending their turn after the sequence of moves, is calculated.

Then the neural network is used to estimate the number of dice that player will lose until the start of their next turn.

Taking into account the probability of success of each move (and also evaluating what happens in the failure states) an estimate of how many dice that player will have at the start of their next turn is constructed for each sequence by subtracting the estimated losses from the calculated gains. The AI player then chooses the move at the start of the sequence with the maximal estimate.

In testing, this approach significantly outperformed the AI from update 2.0.