6 Deep RL in practice

6.1 Limitations

Excellent blog post from Alex Irpan on the limitations of deep RL: https://www.alexirpan.com/2018/02/14/rl-hard.html

Another documented critic on deep RL: https://thegradient.pub/why-rl-is-flawed/

6.2 Simulation environments

Standard RL environments are needed to better compare the performance of RL algorithms. Below is a list of the most popular ones.

import gym
env = gym.make("Taxi-v1")
observation = env.reset()
for _ in range(1000):
    action = env.action_space.sample()
    observation, reward, done, info = env.step(action)

6.3 Algorithm implementations

State-of-the-art algorithms in deep RL are already implemented and freely available on the internet. Below is a preliminary list of the most popular ones. Most of them rely on tensorflow or keras for training the neural networks and interact directly with gym-like interfaces.