Control-oriented Energy-Based Actionable World Model for Decision-Making and Process Control

TMLR Paper7119 Authors

Abstract:

We introduce the \emph{Energy-Based Actionable World Model} (EBAWM), a hybrid
world-modeling framework for industrial process forecasting and control that
combines deterministic state-space dynamics with an energy-based transition
critic. EBAWM is designed for long-horizon, high-stakes decision-making, where
reliable recursive prediction requires both stable state evolution and
principled uncertainty awareness. In contrast to modern deep time-series models—such
as CNNs, RNNs, and Transformers—that operate primarily as input--output predictors, EBAWM
maintains an explicit, recursively propagated state tied to physically
meaningful system variables. This structure enables state correction,
long-horizon simulation, and direct integration with Receding Horizon Control,
model predictive control, and model-based reinforcement learning.The deterministic transition
model provides a strong inductive bias for system
identification by favoring explicit, Markovian, action-conditioned state
transitions, thereby mitigating representation collapse, a common failure mode
in energy-based learning. Uncertainty is captured through an energy function that evaluates the
plausibility of action-conditioned state transitions, rather than by injecting
stochasticity into the dynamics or relying on model ensembles. High-energy
regions naturally indicate dynamically inconsistent or out-of-distribution
behavior, yielding an interpretable uncertainty-aware signal without assuming a
parametric noise model. Our contributions are: (i) we show that the geometry of
the learned energy landscape encodes
dynamical structure and stability-related properties, enabling
uncertainty-aware forecasting and implicit control;(ii) we introduce a
control-oriented world model that combines recursive,
action-conditioned physical state propagation with energy-based transition
evaluation, supporting online optimization and closed-loop decision-making;
and (iii) we propose a simple and stable energy-based modeling design that avoids
representation collapse by operating on a latent space shaped by a
deterministic forecaster.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Sebastian_Trimpe1

Submission Number: 7119