New framework for weighted programming with countable nondeterminism
Researchers extended weighted programming to include user-defined additive rewards and countable nondeterminism.
Researchers extended the weighted programming framework by Batz et al. They added countable nondeterminism and user-defined additive rewards.
Weighted programming describes mathematical models or optimization problems as imperative programs. These programs are parametrized over semirings.
The new framework uses modules instead of semirings. This allows reasoning about expected resource consumption in probabilistic programs and optimal scores in combinatorial games.
A technical challenge exists because wp-calculi can be unsound over arbitrary modules. This happens because continuity properties break in the presence of countable nondeterminism.
The researchers studied conditions for a sound wp-style transformer. They used weighted decision processes (WDPs), which are transition systems that generalize Markov decision processes.
The team mechanized the framework and soundness results in the interactive theorem prover Lean. This yielded a mechanized proof of a fixed point characterization for minimal expected rewards in Markov decision processes with countable nondeterminism.
Key facts
- The framework extends weighted programming with countable nondeterminism and user-defined additive rewards.
- Programs are now parametrized over modules instead of semirings.
- The framework was mechanized in the interactive theorem prover Lean.
- The work provides a mechanized proof of a fixed point characterization for minimal expected rewards in Markov decision processes with countable nondeterminism.