Implementation of some popular reinforcement learning models
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:
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.
The project uses
python 3.6 and
It can be installed directly from pip like
pip install rlmodels
Below is a summary of how the program works. To see the full documentation click here
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(,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
dqn.plot() #plot reward traces dqn.play(n=200) #observe the agent play
example folder for an analogous use of the other algorithms.
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
Release history Release notifications | RSS feed
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size rlmodels-1.0.8.tar.gz (14.0 kB)||File type Source||Python version None||Upload date||Hashes View|