RlLib is a library for reinforcement learning using Axon. It is based on libraries such as chainerrl and pfrl.
Please note that the API is subject to change as it is under development.
Supported Agents: DQN
Supported Explorers: epsilon_greedy
Currently, only DQN is supported. The goal is to increase the number of supported agents in the future.
RlLib is not yet registered on HexDocs, please use it from Git.
def deps do
[
{:rl_lib, git: "git@github.com:masahiro-999/rl_lib.git"},
]
end## How to use
Still in progress.
Please refer to the CartPole sample for reference.
https://github.com/masahiro-999/cart_pole_rl
defmodule CartPoleRl do
alias RlLib.Agents.DQN
alias RlLib.QNet
def run() do
q_function =
Axon.input("x", shape: {nil, 4})
|> Axon.dense(128, activation: :relu)
|> Axon.dense(128, activation: :relu)
|> Axon.dense(2)
env = EnvGym.new("CartPole-v1", render_mode: "human")
explorer =
RlLib.Explorers.EpsilonGreedyExponentialDecay.new(0.3, 0.3, 200, 2)
agent =
DQN.new(
q_function: q_function,
optimizer: Axon.Optimizers.adam(0.001),
explorer: explorer,
target_update_interval: 20,
buffer_size: 10000,
gamma: 0.9,
batch_size: 100
)
agent = RlLib.Experiments.train_agent(agent, env, 300, "cart-pole-qnet-300",
step_hooks: {&step_hooks/2, nil}
)
IO.puts("----eval----")
RlLib.Experiments.eval_agent(agent, env, 20,
step_hooks: {&step_hooks/2, nil}
)
end
def step_hooks(nil, experiments) do
episode_count = experiments.episode_count
total_reward = experiments.total_reward
agent = experiments.agent
IO.puts("#{episode_count}, #{total_reward} #{inspect(QNet.get_info(agent.q_net))}")
nil
end
end