v0.1.0 — Initial Alpha Release

Changelog

All notable changes to this project will be documented in this file.

The format follows Keep a Changelog.
This project adheres to Semantic Versioning.

---

0.1.0 – 2026-04-28

Initial alpha release. All crates are published together at the same version.

rlevo-core

Added

• State<D> / Observation<D> traits for typed, const-generic environment state and agent perception.
• Action<D> trait hierarchy (DiscreteAction, ContinuousAction, MultiDiscreteAction, MultiBinaryAction) for compile-time action-space safety.
• Environment<D, SD, AD> trait with reset / step / render contract; Snapshot<D> trait with SnapshotBase<D, O, R> concrete type.
• Reward trait with ScalarReward and VectorReward implementations.
• TensorConvertible<B, D> bridge trait for lifting state/action types onto Burn tensors.
• Agent and BenchableAgent traits for uniform agent interaction.
• FitnessEvaluable and Landscape traits for benchmarking evolutionary algorithms.
• BenchEnv, BenchError, BenchStep, Metric, MetricsProvider, and SeedStream (moved from rlevo-benchmarks per ADR-0004).
• util::seed — deterministic SeedStream for reproducible multi-run experiments.
• EnvironmentError and StateError error types with thiserror derives.

rlevo-environments

Added

• Classic control — CartPole, MountainCar, MountainCarContinuous, Pendulum, Acrobot.
• Bandits — KArmedBandit<K>, ContextualBandit, NonStationaryBandit, AdversarialBandit.
• Toy text — Blackjack, CliffWalking, FrozenLake, Taxi.
• Gridworlds (MiniGrid-style) — Empty, DoorKey, Memory, FourRooms, Crossing, LavaGap, MultiRoom, Unlock, UnlockPickup, GoToDoor, DistShift, DynamicObstacles; shared GridCore (grid, entity, action, direction, observation, render, reward, dynamics).
• Box2D physics (box2d feature, rapier2d) — BipedalWalker, LunarLander (discrete and continuous action spaces), CarRacing.
• Locomotion (locomotion feature, rapier3d) — Reacher, Swimmer, InvertedPendulum, InvertedDoublePendulum.
• Games — Chess (full move generation and board state), ConnectFour.
• Optimisation landscapes — Sphere, Ackley, Rastrigin for benchmarking evolutionary algorithms.
• Wrappers — TimeLimit wraps any Environment with an episode step cap.
• Bench adapter (bench feature) — BenchAdapter and preset Suite factories to drive any environment from rlevo-benchmarks.
• ASCII render backend for text-based environments.

rlevo-evolution

Added

• Strategy<B> pure trait (init / ask / tell / best) — stateless, parallelism-friendly, trivially checkpointable.
• EvolutionaryHarness<B, S, F> — wraps any Strategy as a BenchEnv.
• BatchFitnessFn trait with FromFitnessEvaluable adapter.
• GenomeKind enum (RealValued, Binary, Integer, Program).
• Classical families — GeneticAlgorithm (real-valued, SBX crossover + polynomial mutation), BinaryGeneticAlgorithm (one-point/uniform crossover + bit-flip mutation), EvolutionStrategy ((1+1), (1+λ), (μ,λ), (μ+λ) with self-adaptive σ), EvolutionaryProgramming (Gaussian perturbation + tournament), DifferentialEvolution (Rand/1/Bin, Best/1/Bin, CurrentToBest/1/Bin), CartesianGeneticProgramming (symbolic regression via CGP graph).
• Metaheuristics — ParticleSwarmOptimization, AntColonyOptimizationReal, AntColonyOptimizationPermutation, ArtificialBeeColony, FireflyAlgorithm, BatAlgorithm, CuckooSearch (Lévy flights via Mantegna), GreyWolfOptimizer, SalpSwarmAlgorithm, WhaleOptimizationAlgorithm.
• Genetic operators (ops) — selection (tournament, roulette, rank, SUS, elitism, NSGA-II crowding), crossover (uniform, one-point, multi-point, SBX, BLX-α, intermediate), mutation (Gaussian, uniform, polynomial, bit-flip, inversion), replacement (generational, steady-state, elitist, comma, plus).
• Custom CubeCL kernels (custom-kernels feature) — fused pairwise-attract (Firefly large-N path) and fused Lévy-flight (Cuckoo/Bat) kernels; pure-tensor fallbacks used when feature is off.
• PopulationState tensor wrapper; ShapingFn fitness shaping (linear rank, exponential rank, truncation).

rlevo-reinforcement-learning

Added

• Replay memory — PrioritizedExperienceReplay (uniform-sampling mode in v0.1.0); TrainingBatch typed container.
• Experience — ExperienceTuple (s, a, r, s', done), History trajectory buffer.
• Metrics — AgentStats (per-step), PerformanceRecord (per-episode).
• DQN — DqnModel, DqnAgent, DqnTrainingConfig; ε-greedy exploration schedule; Double-DQN target option.
• C51 — C51Model, C51Agent, C51TrainingConfig; Bellman projection onto N-atom support; categorical cross-entropy loss.
• QR-DQN — QrDqnModel, QrDqnAgent, QrDqnTrainingConfig; quantile Huber loss; no [v_min, v_max] required.
• PPO — PpoAgent, PpoTrainingConfig, RolloutBuffer, GAE advantages, CategoricalPolicyHead (discrete), TanhGaussianPolicyHead (continuous), clipped surrogate + value loss, early-stop on approx_kl.
• PPG — PpgAgent, PpgConfig, AuxBuffer, PpgCategoricalPolicyHead; interleaved policy-phase + auxiliary-phase with KL distillation.
• DDPG — DdpgAgent, DdpgTrainingConfig; deterministic actor + Q-critic; Polyak target sync; Gaussian exploration noise.
• TD3 — Td3Agent, Td3TrainingConfig; twin-critic min-bootstrap; delayed actor updates; target-policy smoothing.
• SAC — SacAgent, SacTrainingConfig; squashed-Gaussian stochastic actor; twin critics; learnable temperature α with auto-tuning toward -|A|.
• Shared EpsilonGreedy schedule (DQN / C51 / QR-DQN) and GaussianNoise exploration (DDPG / TD3).

rlevo-benchmarks

Added

• Evaluator — drives any BenchEnv for N episodes, collecting per-step and per-episode metrics.
• Suite — ordered sequence of (env, evaluator) pairs with shared reporter.
• Metrics — EaMetrics (best fitness, population diversity, convergence rate), RlMetrics (episode return, episode length, sample efficiency).
• Reporters — JsonReporter (newline-delimited JSON), LoggingReporter (tracing spans), TuiReporter (ratatui live dashboard, tui feature).
• Checkpoint and Storage traits for saving/resuming benchmark state.
• rayon-parallel episode evaluation for multi-seed sweeps.

rlevo-hybrid

Added

• Stub crate establishing the dependency wiring between rlevo-evolution and rlevo-reinforcement-learning. No hybrid strategies are implemented in v0.1.0; see the crate README for the v0.2.0 roadmap.

rlevo (umbrella)

Added

• Re-exports all public APIs from every workspace crate behind a single rlevo entry point.
• keywords: reinforcement-learning, evolutionary, deep-learning, burn, neural-network.
• categories: science, algorithms, simulation.
• Full example suite (35 examples across gridworlds, classic control, Box2D, locomotion, evolutionary showcases, RL algorithms, and benchmarks harness).
• Cross-crate integration tests.

---