Compete and Compose: Learning Independent Mechanisms for Modular World Models

arXiv 24.4 2024 51.3 method

TLDR

COMET learns reusable independent mechanisms via competition and composition for modular world models, enabling efficient adaptation.

Reasoning

The paper introduces a novel two-step training process (competition and composition) that encourages emergence of independent mechanisms, improving sample efficiency and interpretability. However, it is only evaluated on image-based observations without real-world validation, and scalability to complex environments is unclear.

Read-first score

Read-first score 51.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 8%

75.1

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

Topical relevance 42%

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

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=none

Field roles

Candidate

Rank sensitivity

Stability: volatile; rank range: 389.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

video world model

interactive world model

Deep Analysis

Innovations

Winner-takes-all gradient allocation to encourage emergence of independent mechanisms
Two-step competition and composition process for learning and reusing modular mechanisms
Explicit reuse of prior knowledge enabling efficient and interpretable adaptation to new environments

Methodology

COMET is a modular world model trained on multiple environments with varying dynamics. It uses a two-step process: a competition phase with winner-takes-all gradient allocation to encourage independent mechanisms, followed by a composition phase where the model learns to re-compose these mechanisms to capture dynamics of intervened environments.

Key Results

COMET captures recognisable mechanisms without supervision and adapts to new environments with varying numbers of objects with improved sample efficiency compared to conventional finetuning approaches.

Resources

Paper