TEP-Dreamer Implementation Skill

To add a safety shield to a DreamerV3 reinforcement learning pipeline, you implement a quadratic optimization-based action correction mechanism that minimally alters RL actions to satisfy defined safety budgets under a two-tier constraint scheme.

Quantile regression in world-model reinforcement learning safety is used within a Constraint Head to predict physical constraint margins with calibrated uncertainty, which effectively guides safe interventions and risk-bounded action corrections.

To evaluate reinforcement learning generalization on the Tennessee Eastman Process, you configure 4-axis holdout tests that assess policy robustness across disturbance type, intensity, transition, and compound disturbances.

Yes, you can use cvxpy and osqp to build a risk-bounded safety shield for DreamerV3, as these dependencies provide the required quadratic and convex optimization capabilities for action correction.

Yes, you need PyTorch and gymnasium to train a safe DreamerV3 world model, as PyTorch handles latent dynamics learning and imagined rollouts while gymnasium provides the Tennessee Eastman Process environment interface.

A limitation of using DreamerV3 for Tennessee Eastman Process control is the necessity of a two-tier safety constraint scheme to handle Tier-1 trips and Tier-2 alarms, requiring complex convex optimization to maintain safety budgets.