Crate rlevo_core 

Core types and traits for evolutionary deep reinforcement learning.

rlevo-core defines the shared vocabulary used across the entire rlevo workspace. Every other crate — rlevo-reinforcement-learning, rlevo-evolution, rlevo-environments, rlevo-benchmarks — depends on these primitives. No concrete algorithms or environments live here.

Module map

Module	What it provides
base	Reward, Observation, State, Action, TensorConvertible, UpdateFunction — the primitive trait vocabulary
action	DiscreteAction, MultiDiscreteAction, ContinuousAction — layered action-space extensions
state	MarkovState, BeliefState, HiddenState, LatentState, StateAggregation — POMDP and latent-space extensions
environment	Environment, Snapshot, SnapshotBase, EpisodeStatus, EnvironmentError — the agent/environment protocol
reward	ScalarReward — the standard single-value reward concrete type
evaluation	BenchEnv, BenchStep, BenchError — object-safe environment interface for harnesses
fitness	BenchableAgent, FitnessEvaluable, Landscape, Metric, MetricsProvider — inference-only agent and fitness evaluation
render	AsciiRenderable, Renderer, styled/palette/payload sub-modules — optional debug and TUI visualization layer
agent	Reserved; empty in v0.1.x while the unified agent trait hierarchy stabilizes
util	Shared utility helpers

Const-generic RANK

Several traits — Observation, State, Action — are parameterized by a const generic R (rank) that denotes the number of tensor axes (equivalent to NumPy’s ndim or Burn’s Tensor<B, R>). This encodes shape compatibility at compile time: a rank-1 observation and a rank-2 observation cannot be mixed up without a compile error.

Environment<R, SR, AR>
  ├── StateType     : State<SR>        (shape: [usize; SR])
  ├── ObservationType: Observation<R>  (shape: [usize; R])
  ├── ActionType    : Action<AR>       (shape: [usize; AR])
  └── SnapshotType  : Snapshot<R, ...>

Episode loop sketch

The basic agent/environment interaction loop follows this pattern:

env.reset() → Snapshot { observation, reward, status: Running }
  loop:
    agent selects action from observation
    env.step(action) → Snapshot { observation, reward, status }
    break when status.is_done()

EpisodeStatus::Terminated and EpisodeStatus::Truncated are kept distinct so RL algorithms can bootstrap value correctly at truncation boundaries.

Modules

action

Layered action-space traits: DiscreteAction, MultiDiscreteAction, and ContinuousAction.

agent

Reserved for a future unified agent trait hierarchy.

base

Primitive trait vocabulary: Reward, Observation, State, Action, TensorConvertible, and UpdateFunction.

environment

Agent/environment interaction protocol.

evaluation

Object-safe environment interface for benchmarking harnesses.

fitness

Inference-only agent and fitness-evaluation traits.

render

Optional rendering layer for debug output and TUI visualization.

reward

Concrete reward types.

state

Advanced state abstractions for POMDPs and latent representations.

util

Shared utility helpers used across rlevo-core modules. Shared utilities used across rlevo-core consumers.