ReinforcementLearning.jl Template

]add ReinforcementLearningBase 

]add ReinforcementLearningCore ReinforcementLearningEnvironments ReinforcementLearningZoo Flux StatsBase

]st

using ReinforcementLearningBase,ReinforcementLearningCore, ReinforcementLearningEnvironments, ReinforcementLearningZoo

using Flux

using StatsBase

env = CartPoleEnv(; T = Float32, seed = 11)

ns, na = length(rand(get_observation_space(env))), length(get_action_space(env))

agent = Agent(

  policy = QBasedPolicy(

    learner = BasicDQNLearner(

      approximator = NeuralNetworkApproximator(

        model = Chain(

          Dense(ns, 128, relu; initW = seed_glorot_uniform(seed = 17)),

          Dense(128, 128, relu; initW = seed_glorot_uniform(seed = 23)),

          Dense(128, na; initW = seed_glorot_uniform(seed = 39)),

          ) |> gpu,

        optimizer = ADAM(),

        ),

      batch_size = 32,

      min_replay_history = 100,

      loss_func = huber_loss,

      seed = 22,

      ),

    explorer = EpsilonGreedyExplorer(

      kind = :exp,

      ϵ_stable = 0.01,

      decay_steps = 500,

      seed = 33,

      ),

    ),

  trajectory = CircularCompactSARTSATrajectory(

    capacity = 1000,

    state_type = Float32,

    state_size = (ns,),

    ),

)

hook = ComposedHook(TotalRewardPerEpisode(), TimePerStep())

run(agent, env, StopAfterStep(10000), hook)

@info "stats for BasicDQNLearner" avg_reward = mean(hook[1].rewards) avg_fps =

1 / mean(hook[2].times)