Train a neural dynamical surrogate not only to reproduce measured trajectories, but also to reproduce the plant's economically optimal decision and objective value. Add a differentiable decision loss obtained by solving the surrogate's inner optimization problem, and reject models that fit observations while producing extra local optima or a shifted optimum.
Replace pointwise differential PINN residuals with integral residuals tested against smooth functions, so the network can represent shocks without requiring derivatives of a discontinuous prediction. Add a one-sided entropy penalty to select the physically admissible weak solution rather than an arbitrary shock or rarefaction solution.
Replace a single value critic with graph-indexed critics V_alpha and graph-indexed policy heads pi_A, where the labeled graph covers every possible environment mode at every step. Train sampled Bellman inequalities rather than only equality-based temporal-difference errors, and select the policy head using the paper's min-max reachability rule. This targets robust RL settings in which the transition mode can change arbitrarily, such as unknown actuator regimes, domain randomization modes, or…
When an intervention variable is sampled from a distribution rather than fixed to a point, train a predictor on the full joint distribution instead of replacing intervention with conditioning on its realized value. This prevents selection bias caused by conditioning on mediators or downstream observations that reveal information about the random intervention.
Use the reachable-safe-set viewpoint to make training data generation adaptive: maintain an approximation of the states reached by the current neural policy, identify boundary regions with weak barrier margin, and sample there until the set is sufficiently covered. This replaces random rollout expansion with a measurable coverage condition that can support finite-sample safety claims.
Attach an adaptive conformal error radius to every predicted agent and forecast horizon, then use that radius to inflate collision constraints or mask unsafe actions in a learned policy. Unlike a fixed heuristic margin, the radius automatically grows after systematic prediction failures and shrinks when the predictor is accurate, providing an explicit accuracy-versus-conservatism control.
Train a neural controller as a uniformly accurate surrogate of a trusted but expensive controller, and use a measured small-gain condition to decide whether the surrogate is safe for closed-loop deployment. The approximation tolerance becomes an interpretable residual-state budget instead of an opaque validation metric.
Replace a single Gaussian, Laplace, or Huber residual model with a conditional mixture containing an inlier component, a body component, and an explicit generalized-Pareto tail. The network learns both the prediction and the probability that an error belongs to the extreme tail, allowing rare large errors to be modeled without making the entire loss excessively sensitive to ordinary noise.