PG-DPO is a new variational framework that replaces Bellman recursion with a Pontryagin-guided adjoint-MC projection for RL under non-exponential discounting and shows gains on hyperbolic and survival benchmarks.
On the Well-posedness of Hamilton-Jacobi-Bellman Equations of the Equilibrium Type
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
This paper studies the well-posedness of a class of nonlocal parabolic partial differential equations (PDEs), or equivalently equilibrium Hamilton-Jacobi-Bellman equations, which has a strong tie with the characterization of the equilibrium strategies and the associated value functions for time-inconsistent stochastic control problems. Specifically, we consider nonlocality in both time and space, which allows for modelling of the stochastic control problems with initial-time-and-state dependent objective functionals. We leverage the method of continuity to show the global well-posedness within our proposed Banach space with our established Schauder prior estimate for the linearized nonlocal PDE. Then, we adopt a linearization method and Banach's fixed point arguments to show the local well-posedness of the nonlocal fully nonlinear case, while the global well-posedness is attainable provided that a very sharp a-priori estimate is available. On top of the well-posedness results, we also provide a probabilistic representation of the solutions to the nonlocal fully nonlinear PDEs and an estimate on the difference between the value functions of sophisticated and na\"{i}ve controllers. Finally, we give a financial example of time inconsistency that is proven to be globally solvable.
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Beyond the Bellman Recursion: A Pontryagin-Guided Framework for Non-Exponential Discounting
PG-DPO is a new variational framework that replaces Bellman recursion with a Pontryagin-guided adjoint-MC projection for RL under non-exponential discounting and shows gains on hyperbolic and survival benchmarks.