World Models [draft]
World Models Learning the transition of the environment Reference: world models paper world models author's blog The goal of World Model is to learn a transition model which is able to characterize the probability distribution $p(s_{t}|s_{t-1}, a_{t-1})$. Such pre-knowledge of the environment transition convert the tasks into model based, thereby enabling the application of advanced methodologies, including planning and model-based reinforcement learning (MBRL). In this paper, the World Model is built with two distinct yet interconnected components: Vision Model (V) and Memory Model (M) where V learns to encode the high-dimensional observation into low-dimensional latent space and M learns to predict low-dimensional representation of the future observation....