torchagent 0.2.4

Deep-Q learning with pytorch

Welcome to the torchagent repository. This repository contains the sources for the torchagent library.

Contents

• What is it?

• Which algorithms are included?

• Installation

• Usage

What is it?

torchagent is a library that implements various reinforcement learning algorithms for PyTorch. You can use this library in combination with openAI Gym to implement reinforcement learning solutions.

Which algorithms are included?

Currently the following algorithms are implemented:

• Deep Q Learning

• Double Q Learning

Installation

You can install the library using the following command:

pip install torchagent

Usage

The following code shows a basic agent that uses Deep Q Learning.

from torchagent.memory import SequentialMemory
from torchagent.agents import DQNAgent

import torch
import torch.nn as nn
import torch.optim as optim

class PolicyNetwork(nn.Module):
    def __init__(self):
        self.linear = nn.Linear(210 * 160, 3)

    def forward(self, x):
        return self.linear(x)

policy_network = PolicyNetwork()
memory = SequentialMemory(20)
agent = DQNAgent(2, policy_network, nn.MSELoss(), optim.Adam(policy_network.parameters()), memory)

env = gym.make('Assault-v0')

for _ in range(50):
    state = env.reset()

    for t in count():
        action = agent.act(state)
        next_state, reward, done, _ = env.step(agent.act(state))

        agent.record(state, action, next_state, reward, done)
        agent.train()

        state = next_state

        if done:
            break