MuZero
Lets define our alphabet soup: * h - calculates latent representation of state s from the past observation (board state, or previous frames) * s - state, latent representation of the environment * f - calculates p(policy) and v(value function) from s(state) * p - policy value for each action * v - value function, based on the reward. For atari n-step reward, final reward for board games. * a - some action, sampled from π/p when interacting with the environment, sampled from replay buffer during training. * g - calculates next s(state) and immediate reward(r), recieves previous state and an action as input * r - immediate reward * π - policy, approximately p
