Finite Discounted MDPs

This module gives TorchLean a small, proof-friendly finite-state discounted MDP layer.

Design choices:

• transitions are deterministic and total,
• the latent state space is Fin nStates,
• the action space is Fin nActions,
• Bellman operators are defined directly on typed value tensors.

This first formalization supports the core objects used by RL theory: policies, value functions, state-action values, and Bellman operators.

References:

• Bellman, Dynamic Programming (1957)
• Puterman, Markov Decision Processes (1994)
• Sutton and Barto, Reinforcement Learning: An Introduction
• Gymnasium and TorchRL are useful runtime reference points, while this file defines the pure finite-MDP semantics rather than modeling replay buffers or collectors.

Naming note:

• In this namespace, FiniteMDP, Policy, ValueFunction, and the Bellman operators refer to deterministic finite tensor MDPs.
• Spec.RL.FiniteStochastic and Spec.RL.Markov deliberately reuse standard RL words such as MDP and Policy inside their own namespaces. We keep the short mathematical names there because the fully qualified names already say which semantic layer is being used.

@[reducible, inline]

abbrev Spec.RL.ValueFunction (α : Type) [TorchLean.Storage α] (nStates : ℕ) : Type

Value function over a finite state space.

@[reducible, inline]

abbrev Spec.RL.Policy (nStates nActions : ℕ) : Type

Deterministic policy over a finite state / action space.

structure Spec.RL.FiniteMDP (α : Type) (nStates nActions : ℕ) : Type

Finite discounted MDP with deterministic transitions.

• initialState : Fin nStates

  Canonical reset state.

• step : Fin nStates → Fin nActions → StepResult (Fin nStates) α

  One-step deterministic transition / reward dynamics.

• discount : α

  Discount factor used by Bellman operators.

def Spec.RL.FiniteMDP.toEnv {α : Type} {nStates nActions : ℕ} (mdp : FiniteMDP α nStates nActions) : Env (Fin nStates) (Fin nActions) (Fin nStates) α

View a finite MDP as a Gym-style environment with observations equal to latent states.

def Spec.RL.valueAt {α : Type} [TorchLean.Storage α] {nStates : ℕ} (values : ValueFunction α nStates) (state : Fin nStates) : α

Lookup a state's value.

def Spec.RL.stateActionValue {α : Type} [TorchLean.Storage α] {nStates nActions : ℕ} [Zero α] [One α] [Add α] [Mul α] (mdp : FiniteMDP α nStates nActions) (values : ValueFunction α nStates) (state : Fin nStates) (action : Fin nActions) : α

One-step state-action value induced by a candidate value function.

def Spec.RL.actionValues {α : Type} [TorchLean.Storage α] {nStates nActions : ℕ} [Zero α] [One α] [Add α] [Mul α] (mdp : FiniteMDP α nStates nActions) (values : ValueFunction α nStates) (state : Fin nStates) : TorchLean.Tensor α [nActions]

All state-action values $Q_v(s,\mathord{\cdot})$ for a fixed state and candidate value function.

def Spec.RL.bellmanPolicy {α : Type} [TorchLean.Storage α] {nStates nActions : ℕ} [Zero α] [One α] [Add α] [Mul α] (mdp : FiniteMDP α nStates nActions) (policy : Policy nStates nActions) (values : ValueFunction α nStates) : ValueFunction α nStates

Bellman operator for a deterministic policy.

def Spec.RL.bellmanOptimality {α : Type} [TorchLean.Storage α] {nStates nActions : ℕ} [Zero α] [One α] [Add α] [Mul α] [LinearOrder α] [Fact (0 < nActions)] (mdp : FiniteMDP α nStates nActions) (values : ValueFunction α nStates) : ValueFunction α nStates

Bellman optimality operator for a finite action space.