rlmodels: a reinforcement learning library

This project is a collection of some popular optimisation algorithms for reinforcement learning problem. At the moment the available models are:

• DQN
• DDPG
• CMAES
• AC

with some more going to be added in the future.

It works with Pytorch models and environment classes like the OpenAI gym ones. Any environment class wrapper that mimic their basic functionality should be fine, but more on that below.

Getting Started

Prerequisites

The project uses python >= 3.6, torch 1.1.0 and relies on OpenAI's gym environments.

Installing

It can be installed directly from pip like

pip install rlmodels

Usage

Below is a summary of how the program works. To see the full documentation click here

Initialization

The following is an example with the popular CartPole environment using a double Q network. First the setup.

import numpy as np
import torch
import torch.optim as optim
import gym

from rlmodels.models.DQN import *
from rlmodels.nets import FullyConnected

import logging

#logger parameters
FORMAT = '%(asctime)-15s: %(message)s'
logging.basicConfig(level=logging.INFO,format=FORMAT,filename="model_fit.log",filemode="a")

max_ep_ts = 200

env = gym.make('CartPole-v0')
env._max_episode_steps = max_ep_ts

env.seed(1)
np.random.seed(1)
torch.manual_seed(1)

the episode and timepstep numbers as well as the average reward trace is logged to the file model_fit.log. Setting the logging level to logging.DEBUG will also log information about gradient descent steps.

The library also has a basic network definition, FullyConnected, to which we only need to specify number and size of hidden layer, input and output sizes, and last activation function. It uses ReLU everywhere else by default.

let's create the basic objects

dqn_scheduler = DQNScheduler(
	batch_size = lambda t: 200, #constant
	exploration_rate = lambda t: max(0.01,0.05 - 0.01*int(t/2500)), #decrease exploration down to 1% after 10,000 steps
	PER_alpha = lambda t: 1, #constant
	PER_beta = lambda t: 1, #constant
	tau = lambda t: 100, #constant
	agent_lr_scheduler_fn = lambda t: 1.25**(-int(t/1000)), #decrease step size every 2,500 steps,
	steps_per_update = lambda t: 1) #constant

agent_lr = 0.5 #initial learning rate
agent_model = FullyConnected([60],4,2,None)
agent_opt = optim.SGD(agent_model.parameters(),lr=agent_lr,weight_decay = 0, momentum = 0)

agent = Agent(agent_model,agent_opt)

the models take a scheduler object as argument which allows parameters to be changed at runtime accordint to user-defined rules. For example reducing learning rate and exploration rate after a certain number of iterations, as above. Finally, all gradient-based algorithms receive as input an Agent instance that contains the network deffinition and optimisation algorithm. Once all this is setup we are good to go.

dqn = DQN(agent,env,dqn_scheduler)

dqn.fit(
	n_episodes=170,
	max_ts_by_episode=max_ep_ts,
	max_memory_size=2000,
	td_steps=1)

Once the agent is trained we can visualize the reward trace. If we are using an environment with a render method (like OpenAI ones) we can also visualise the trained agent. We can also use the trained model using the forward method of the ddq object or simply extract it with ddq.agent

dqn.plot() #plot reward traces
dqn.play(n=200) #observe the agent play

see the example folder for an analogous use of the other algorithms.

Environment

For custom environments or custom rewards, its possible to make a wrapper tha mimics te behavior of the step() and reset() function of gym's environemnts

class MyCustomEnv(object):
	def __init__(self,env):
		self.env = env
	def step(self,action):
		## get next state s, reward, termination flag (boolean) and any additional info
		return s,r, terminated, info #need to output these 4 things (info can be None)
	def reset(self):
		#something
	def seed(self):
		#something