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 the raw HJB residual loss of a neural PDE solver with a parametrix-preconditioned fixed-point target. At each local space-time patch, analytically propagate terminal values and source terms through a Gaussian kernel whose covariance uses a frozen diffusion matrix, while asking the network to learn only the variable-coefficient correction. This should reduce the burden on the network to represent stiff high-frequency diffusion dynamics and improve short-horizon convergence.
Train an input-generation policy or differentiable signal parameterization to produce trajectories that cover the joint input-state feature space while remaining informative for every plausible neural world model. Replace single-model experiment design by an expectation over an ensemble of models, and optimize this objective with stochastic model and trajectory samples.
Retain the iteration at which each state enters each modal winning set and use that integer as a dense training target for a neural critic. The policy is additionally encouraged to choose transitions that decrease every finite modal distance, supplying progress information even when the environment reward is sparse.