ReinforcementLearning.jlのサンプル

ReinforcementLearning_sample_CartPoleEnv.jl

using Pkg; Pkg.add.(["ReinforcementLearning", "ReinforcementLearningEnvironments", "Flux", "StatsBase", "Plots"])
# import Random # hide
# Random.seed!(1) # hide

using ReinforcementLearning, ReinforcementLearningEnvironments, Flux
# using StatsBase:mean

env = CartPoleEnv(;T=Float32)
ns, na = length(observation_space(env)), length(action_space(env))  # (4, 2)

device = :cpu


layer1 = Dense(ns, 128, relu)
layer2 = Dense(128, 128, relu)
layer3 = Dense(128, na)
neural_network_q = NeuralNetworkQ(model = Chain(layer1, layer2, layer3),
                                  optimizer = ADAM(),
                                  device = device)

ϵ_selector = EpsilonGreedySelector{:exp}(ϵ_stable = 0.01, decay_steps = 500)

q_base_policy = QBasedPolicy(learner = BasicDQNLearner(approximator = neural_network_q,
                                                       batch_size = 32,
                                                       min_replay_history = 100,
                                                       loss_fun = huber_loss),
                             selector = ϵ_selector)

circular_risa_buffer = circular_RTSA_buffer(capacity = 1000,
                                            state_eltype = Float32,
                                            state_size = (ns,))
agent = Agent(
    π      = q_base_policy,
    buffer = circular_risa_buffer
)

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

run(agent, env, StopAfterStep(10000; is_show_progress=true); hook = hook)

using Plots
gr()

plot(hook[1].rewards, xlabel="Episode", ylabel="Reward", label="")
savefig("a_quick_example_cartpole_cpu_basic_dqn.png");


for b ∈ agent.buffer
  env.state = b.state
  render(env)
end