What's Going on in a Planning Process?

The architecture is a transformer and we feed into it the board state, the recent action history. And basically what we're using this for is a value function. The way the planning works diplomacy is a simultaneous move game. So on every iteration we have our initial prediction and then we say, okay, well we can reasonably assume that players will choose the value function. We update the probability distribution of our actions to be more biased towards like to put higher probability on the actions that have higher expected value.