world2vec learns mechanistic world models from observations and actions
Original title: world2vec: Mechanistic world models through latent dynamics learning
Latent dynamics use at least ten times fewer parameters than DreamerV3, TD-MPC2 and LeWM, without reward.
The paper introduces world2vec to learn mechanistic world models directly from observations and actions, without reward information. It learns latent dynamics that reproduce the true system over long prediction horizons, with at least one order of magnitude fewer parameters than DreamerV3, TD-MPC2 and LeWM. The authors report state-of-the-art control of continuous environments from pixels and accurate system identification across a benchmark of 71 attractors.
Unconfirmed: State-of-the-art control is the authors' claim; the abstract gives no numerical scores.
- Training uses observations and actions and does not use reward.
- The comparison models named are DreamerV3, TD-MPC2 and LeWM.
- System identification is evaluated on a benchmark of 71 attractors.
- The stated goal is interpretability and scientific discovery, not only control.
From the original article
1 min read at arxiv.orgWorld models can achieve strong control performance without faithfully capturing the dynamics of their environment, limiting their potential for mechanistic interpretability and scientific discovery. Here, we introduce world2vec to learn mechanistic world models without reward information directly from observations and actions.
By Tobias Schmidt · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.