Digital twin speeds up reinforcement learning for lab robots

Dataset for deciding between evidence gathering and action in control-system incidents

The authors present TE-GvA, a counterfactual decision dataset for fault-driven incident response on the revised Tennessee Eastman process: 9,908 decision points from 1,990 simulated incidents, each with 4–35 candidate actions labelled via branched rollouts. With labels computed under the true hidden cause, gathering evidence is the best choice in only 1.1–1.3% of decisions.

Sparse stochastic twins can outperform exact digital twins

The study shows that keeping weakly dependent or auxiliary degrees of freedom raises the dimensionality of learning and control problems, so a sparse stochastic surrogate can beat an exact high-dimensional model. It introduces a method for dependence discovery at the level of the full conditional law, combining conditional generative modeling, Gaussian-process distillation and kernel mode decomposition.

Consolidating dispersed DSO data into a digital twin of the distribution grid