World Models Increase Autonomy in Reinforcement Learning

arXiv 24.8 2024 39.7 method

TLDR

Proposes MoReFree, a model-based RL agent for reset-free tasks, outperforming prior methods with less supervision.

Reasoning

The paper addresses the reset-free RL setting and demonstrates that model-based methods outperform prior state-of-the-art, with a novel agent MoReFree. However, the abstract lacks explicit details on world model architecture or real-world validation, and the term 'world model' is used only in the title without elaboration.

Read-first score

Read-first score 39.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 15.

Recency 8%

75.1

Uses a gentle age decay so recent papers surface without erasing older foundations. 2024

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=github

Topical relevance 42%

21.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 140.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Demonstrating superiority of model-based RL methods over prior state-of-the-art in the reset-free setting
Proposing MoReFree agent with adapted exploration and policy learning mechanisms that prioritize task-relevant states
Achieving data-efficient performance without access to environmental reward or demonstrations, outperforming privileged baselines that require supervision

Methodology

The paper adapts model-based RL to the reset-free setting, then identifies limitations of this direct extension and proposes the MoReFree agent. MoReFree incorporates two key mechanisms—exploration and policy learning—both adapted to prioritize task-relevant states. The method is evaluated on various reset-free tasks without reward or demonstrations, compared to prior state-of-the-art and privileged baselines.

Key Results

MoReFree outperforms all prior state-of-the-art methods and privileged baselines that require supervision, demonstrating superior data-efficiency across various reset-free tasks.

Limitations

Limitations inherent to the direct extension of MBRL to the reset-free setting are identified but not detailed in the abstract

Resources

Paper Project