moonrllab
MoonRLLab is a discrete reinforcement-learning lab for MoonBit.
moonbit
reinforcement-learning
q-learning
sarsa
gridworld
moon add liuzhiyug/moonrllab@0.1.1
Download zipAuthor
Version
0.1.1
License
Apache-2.0
Repository
Last updated
2 days ago
Downloads
5
README#MoonRLLab
MoonRLLab is a reusable tabular reinforcement-learning toolkit for MoonBit.
#Library quick start
import { "liuzhiyug/moonrllab" }
fn main {
let report = @moonrllab.train_with_memory_logger(8, 40, 20260815)
println(report.compact_line())
}The checked executable equivalent is in examples/basic:
moon run examples/basicThe generic Trainer::train works through open Environment, Agent,
Logger and Policy traits. Built-in examples cover GridWorld, CliffWalking,
RandomWalk and a multi-armed bandit. See README.md for the full API and
acceptance evidence.
#Agent
pub(open) trait Agent {
fn reset_episode(Self) -> Unit
fn choose_action(Self, Int) -> Int
fn learn(Self, Transition, Int?) -> Unit
fn epsilon(Self) -> Double
fn q_report(Self, Int) -> String
}#Environment
pub(open) trait Environment {
fn reset(Self) -> Int
fn actions(Self) -> Array[Int]
fn state_space(Self) -> Array[Int]
fn step(Self, Int) -> Transition
fn render(Self) -> String
}#Logger
pub(open) trait Logger {
fn start_episode(Self, Int) -> Unit
fn step(Self, Int, Int, Int, Double, Int, Bool, Int) -> Unit
fn finish_episode(Self, EpisodeRecord) -> Unit
fn finish(Self, TrainingReport) -> Unit
}#Policy
#BanditEnv
A reproducible k-armed bandit. Rewards are bounded and generated from a
deterministic pseudo-random stream, which makes it suitable for CI.
impl Environment for BanditEnv
#BanditPolicy
pub struct BanditPolicy {
values : Array[Double]
counts : Array[Int]
schedule : EpsilonSchedule
rng : LcgRng
} derive(Debug)#BanditPolicy::new
#BanditPolicy::observe
#BenchmarkCase
#BenchmarkResult
Aggregated results for a fixed-seed benchmark.
#BenchmarkResult::new
#BenchmarkResult::record
fn BenchmarkResult::record(self : BenchmarkResult, episode : Int, reward : Double, steps : Int, solved : Bool) -> Unit#BenchmarkResult::reward_stats
#BenchmarkResult::step_stats
#BenchmarkResult::tail_average
#CliffWalkingEnv
pub struct CliffWalkingEnv {
width : Int
height : Int
start_x : Int
start_y : Int
goal_x : Int
goal_y : Int
x : Int
y : Int
_last_stage : EpisodeStage
} derive(Debug)#ConsoleLogger
impl Logger for ConsoleLoggerfn step(self : ConsoleLogger, episode : Int, state : Int, action : Int, reward : Double, next_state : Int, done : Bool, step : Int) -> Unit#DoubleQLearningAgent
pub struct DoubleQLearningAgent {
left : QTable
right : QTable
policy : EpsilonGreedyPolicy
alpha : Double
gamma : Double
updates : Int
} derive(Debug)#DoubleQLearningAgent::best_action
#DoubleQLearningAgent::choose_action
#DoubleQLearningAgent::learn
fn DoubleQLearningAgent::learn(self : DoubleQLearningAgent, transition : Transition, update_left : Bool) -> Unit#DoubleQLearningAgent::new
fn DoubleQLearningAgent::new(states : Array[Int], actions : Array[Int], policy : EpsilonGreedyPolicy, alpha : Double, gamma : Double) -> DoubleQLearningAgent#DoubleQLearningAgent::update_count
#DoubleQLearningAgent::value
#EnvironmentAudit
#EpisodeAccumulator
pub struct EpisodeAccumulator {
reward : Double
steps : Int
last_state : Int
solved : Bool
} derive(Debug)#EpisodeAccumulator::finish
#EpisodeAccumulator::observe
#EpisodeRecord
#EpisodeTrace
A transition trace retained until an episode finishes.
#EpisodeTrace::push
#EpisodeTrace::returns
#EpsilonGreedyPolicy
impl Policy for EpsilonGreedyPolicyfn choose_action(self : EpsilonGreedyPolicy, state : Int, actions : Array[Int], q_values : Array[Double]) -> Int#EpsilonGreedyPolicy::choose_action
fn EpsilonGreedyPolicy::choose_action(self : EpsilonGreedyPolicy, _state : Int, actions : Array[Int], q_values : Array[Double]) -> Int#EpsilonGreedyPolicy::new
#EpsilonSchedule
#EpsilonSchedule::new
#EpsilonSchedule::value
#EpsilonSchedule::values
#EvaluationConfig
#EvaluationConfig::new
#EvaluationConfig::with_window
#EvaluationPoint
A single observation collected during an evaluation run.
#ExpectedSARSAAgent
pub struct ExpectedSARSAAgent {
table : QTable
policy : EpsilonGreedyPolicy
alpha : Double
gamma : Double
} derive(Debug)impl Agent for ExpectedSARSAAgent#ExpectedSARSAAgent::choose_action
#ExpectedSARSAAgent::epsilon
#ExpectedSARSAAgent::learn
fn ExpectedSARSAAgent::learn(self : ExpectedSARSAAgent, transition : Transition, _next_action : Int?) -> Unit#ExpectedSARSAAgent::new
fn ExpectedSARSAAgent::new(states : Array[Int], actions : Array[Int], policy : EpsilonGreedyPolicy, alpha : Double, gamma : Double) -> ExpectedSARSAAgent#ExpectedSARSAAgent::q_report
#ExpectedSARSAAgent::reset_episode
#ExperimentSummary
#GridWorldEnv
pub struct GridWorldEnv {
width : Int
height : Int
start_x : Int
start_y : Int
goal_x : Int
goal_y : Int
x : Int
y : Int
_last_stage : EpisodeStage
} derive(Debug)impl Environment for GridWorldEnv#LearningRateSchedule
#LearningRateSchedule::new
fn LearningRateSchedule::new(initial : Double, minimum : Double, decay : Double) -> LearningRateSchedule#LearningRateSchedule::value
#MemoryLogger
pub struct MemoryLogger {
started : Int
transitions : Int
finished : Int
last_report : TrainingReport?
} derive(Debug)A silent logger for library and CI consumers that need structured results
without writing progress lines to stdout.
impl Logger for MemoryLoggerfn step(self : MemoryLogger, _episode : Int, _state : Int, _action : Int, _reward : Double, _next_state : Int, _done : Bool, _step : Int) -> Unit#MonteCarloAgent
pub struct MonteCarloAgent {
table : QTable
policy : EpsilonGreedyPolicy
alpha : Double
gamma : Double
visits : Array[Int]
} derive(Debug)#MonteCarloAgent::choose_action
#MonteCarloAgent::learn_episode
#MonteCarloAgent::new
fn MonteCarloAgent::new(states : Array[Int], actions : Array[Int], policy : EpsilonGreedyPolicy, alpha : Double, gamma : Double) -> MonteCarloAgent#MonteCarloAgent::q_value
#MonteCarloAgent::visit_count
#PolicyEvaluation
#PolicyEvaluation::value
#QLearningAgent
pub struct QLearningAgent {
table : QTable
policy : EpsilonGreedyPolicy
alpha : Double
gamma : Double
} derive(Debug)impl Agent for QLearningAgent#QLearningAgent::choose_action
#QLearningAgent::learn
fn QLearningAgent::learn(self : QLearningAgent, transition : Transition, _next_action : Int?) -> Unit#QLearningAgent::new
fn QLearningAgent::new(states : Array[Int], actions : Array[Int], policy : EpsilonGreedyPolicy, alpha : Double, gamma : Double) -> QLearningAgent#QLearningAgent::q_report
#QLearningAgent::q_value
#QLearningAgent::reset_episode
#QLearningAgent::set_q_value
fn QLearningAgent::set_q_value(self : QLearningAgent, state : Int, action : Int, value : Double) -> Unit#RandomWalkEnv
A small non-terminal random-walk benchmark used for value-estimation tests.
impl Environment for RandomWalkEnv#ReplayBuffer
#ReplayBuffer::push
#ReportTable
#RunningStats
pub struct RunningStats {
count : Int
mean : Double
m2 : Double
minimum : Double
maximum : Double
} derive(Debug)Online statistics with numerically stable mean and variance updates.
#SARSAAgent
pub struct SARSAAgent {
table : QTable
policy : EpsilonGreedyPolicy
alpha : Double
gamma : Double
} derive(Debug)impl Agent for SARSAAgent#SARSAAgent::learn
#SARSAAgent::new
fn SARSAAgent::new(states : Array[Int], actions : Array[Int], policy : EpsilonGreedyPolicy, alpha : Double, gamma : Double) -> SARSAAgent#Scorecard
#Scorecard::new
#ScorecardSet
#SeedBank
#Trainer
#Trainer::train
fn[Env : Environment, Ag : Agent, Log : Logger] Trainer::train(self : Trainer, env : Env, agent : Ag, logger : Log) -> TrainingReportA trait-driven training entry point. The agent receives the next action as
an optional hint: SARSA uses it, while Q-learning ignores it. This keeps
the episode loop independent of a concrete environment or agent type.
#Trainer::train_q_learning
fn Trainer::train_q_learning(self : Trainer, env : GridWorldEnv, agent : QLearningAgent, logger : ConsoleLogger) -> TrainingReport#Trainer::train_sarsa
fn Trainer::train_sarsa(self : Trainer, env : GridWorldEnv, agent : SARSAAgent, logger : ConsoleLogger) -> TrainingReport#TrainingReport
#TrainingReport::final_epsilon_value
#TrainingReport::reward_at
#TrainingReport::steps_at
#Transition
#Transition::new
fn Transition::new(state : Int, action : Int, reward : Double, next_state : Int, done : Bool, step : Int) -> Transition#ValidationFinding
pub struct ValidationFinding {
rule : String
severity : ValidationSeverity
message : String
} derive(Eq, Debug)#ValidationReport
pub struct ValidationReport {
findings : Array[ValidationFinding]
passed : Int
notices : Int
failures : Int
} derive(Debug)#ValidationReport::add
#ValidationReport::finding_count
#ValueIterationResult
A value-iteration result that can be inspected independently of a learner.
#ValueIterationResult::action
#ValueIterationResult::greedy_path
fn ValueIterationResult::greedy_path(self : ValueIterationResult, start : Int, goal : Int, limit : Int) -> Array[Int]#ValueIterationResult::value
#audit_score
fn audit_score(audit : EnvironmentAudit, expected_states : Int, expected_actions : Int) -> ScorecardSet#compare_results_table
#compare_training_reports
#discounted_returns
#evaluate_grid_policy
fn evaluate_grid_policy(policy : Array[Int], goal : Int, gamma : Double, limit : Int) -> PolicyEvaluation#greedy_action_map
#gridworld_value_iteration
fn gridworld_value_iteration(width : Int, height : Int, goal : Int, gamma : Double, tolerance : Double, limit : Int) -> ValueIterationResultCompute an optimal bounded policy for a rectangular navigation task.
#linear_interpolate
fn linear_interpolate(left : Double, right : Double, ratio : Double) -> Double#merge_reward_series
#policy_agreement
#project_name
fn project_name() -> StringMoonRLLab is a compact reinforcement-learning lab for discrete problems.
The project focuses on a small but extensible tabular stack:
- environments expose a finite state space and legal actions
- policies choose actions with epsilon-greedy exploration
- agents implement Q-learning and SARSA updates
- the trainer produces a readable episode report
#q_table_snapshot
#run_gridworld_benchmark
#run_random_walk_benchmark
#train_cliff_qlearning_generic
#train_gridworld_expected_sarsa_generic
fn train_gridworld_expected_sarsa_generic(episodes : Int, max_steps : Int, seed : Int) -> TrainingReport#train_gridworld_qlearning_generic
#train_gridworld_sarsa_generic
#train_with_memory_logger
#validate_policy
#validate_schedule
MoonRLLab is a discrete reinforcement-learning lab for MoonBit.
moonbitreinforcement-learningq-learningsarsagridworld
0.1.1
Apache-2.0
Updated 2 days ago
5 downloads