Hieros: Hierarchical Imagination on Structured State Space Sequence World Models

TL;DR

Hieros is a hierarchical policy that learns time abstracted world representations and imagines trajectories at multiple time scales in latent space that allows for more efficient training than RNN- based world models and more efficient imagination than Transformer-based world models.

Abstract

One of the biggest challenges to modern deep reinforcement learning (DRL) algorithms is sample efficiency. Many approaches learn a world model in order to train an agent entirely in imagination, eliminating the need for direct environment interaction during training. However, these methods often suffer from either a lack of imagination accuracy, exploration capabilities, or runtime efficiency. We propose Hieros, a hierarchical policy that learns time abstracted world representations and imagines trajectories at multiple time scales in latent space. Hieros uses an S5 layer-based world model, which predicts next world states in parallel during training and iteratively during environment interaction. Due to the special properties of S5 layers, our method can train in parallel and predict next world states iteratively during imagination. This allows for more efficient training than RNN-based world models and more efficient imagination than Transformer-based world models. We show that our approach outperforms the state of the art in terms of mean and median normalized human score on the Atari 100k benchmark, and that our proposed world model is able to predict complex dynamics very accurately. We also show that Hieros displays superior exploration capabilities compared to existing approaches.

Figures (13)

• Figure 1: On the left: Hierarchical subactor structure of Hieros. Each layer of the hierarchy learns its own latent state world model and interacts with the other layers via subgoal proposal. The action outputs of each actor/critic is the subgoal input of the next lower layer. The output of the lowest level actor/critic is the actual action in the real environment. On the right: Training and imagination procedure of the S5WM. Hieros uses a stack of S5 blocks with their architecture shown above.
• Figure 2: Trajectories for Breakout (top) and Frostbite (bottom). For each, the upper frame is the image observed in the environment and the lower frames are the imagined trajectories of the S5WM of the lowest level subactor.
• Figure 3: Extrinsic, subgoal, and novelty rewards per step for Krull (top) and Breakout (bottom) for the lowest level subactor.
• Figure 4: World model losses for the S5WM and RSSM for Krull and Breakout. The S5WM is able to achieve an overall lower world model loss compared to the RSSM for Krull, while those roles are reversed for Breakout.
• Figure 5: Proposed subgoals for Breakout (top row), Frostbite (middle row), and Freeway (bottom row). The left most frame is the original observation from the environment, and the following frames are the proposed subgoals from the higher level actor. For Breakout, the subgoals are only to increase the level score (marked with the red rectangles) and the ball is not simulated at all, while for Frostbite the subgoals guide the actor towards building up the igloo in the upper right part of the image in order to advance to the next level (red rectangles). For Freeway, which also features a single level and sparse rewards, the subgoals are much more meaningful than for Breakout and guide the actor to move across the road (red rectangles).