12-09-2015, 01:51 PM
I have developed a neural network architecture which is ridiculously effective for the specific task of learning long-ish term strategies to accomplish tasks. Arimaa (or any turn-based adversarial game, really) sounds like exactly the sort of thing it *could* do very well.
It basically consists of a recurrent neural network using backpropagation-through-time - and also capable of direct backprop on the outputs of its previous states. A few of the outputs are reserved for making predictions about how well it will be doing in various timeframes, and it gets positive/negative feedback (on ALL of its past decisions within that timeframe) depending on whether it's beaten or failed to beat the predictions it made that many turns ago. And likewise the predictions can be refined with backpropagation every round (against previous states) using the current state as the 'correct' value.
Of course neural networks are not the "refined brute force" approach that Arimaa was designed to be difficult for. Go is even more resistant to refined brute-force search than chess, so I don't really understand the motivation for creating Arimaa, unless it's good intellectual fun to play as, you know, a game for humans.
But, hmm, I think Go would be very amenable to the recurrent-network approach, too.
It basically consists of a recurrent neural network using backpropagation-through-time - and also capable of direct backprop on the outputs of its previous states. A few of the outputs are reserved for making predictions about how well it will be doing in various timeframes, and it gets positive/negative feedback (on ALL of its past decisions within that timeframe) depending on whether it's beaten or failed to beat the predictions it made that many turns ago. And likewise the predictions can be refined with backpropagation every round (against previous states) using the current state as the 'correct' value.
Of course neural networks are not the "refined brute force" approach that Arimaa was designed to be difficult for. Go is even more resistant to refined brute-force search than chess, so I don't really understand the motivation for creating Arimaa, unless it's good intellectual fun to play as, you know, a game for humans.
But, hmm, I think Go would be very amenable to the recurrent-network approach, too.