search

aleju / self-driving-truck

Self-Driving Truck in Euro Truck Simulator 2, trained via Reinforcement Learning
MIT License
389 stars 96 forks source link

Question - Is it possible to apply the A3C RL method to this environment? #5

Open aurotripathy opened 5 years ago

aurotripathy commented 5 years ago

This is quite nice! There are several A3C PyTorch implementations for Atari. Is it possible to do the same with this Truck environment? Thank you.

aleju commented 5 years ago

A3C is usually based on having multiple workers playing in parallel. This implementation here expects a GUI to run the game in and record its state via screenshots. Getting that to work would probably require multiple machines running the model in parallel, including all the communication code. The implementation is currently not geared towards such a scenario, so it would require a significant rewrite of large parts of the code to get these things -- and thus A3C -- to work.

aurotripathy commented 5 years ago

Thank you. If the environment was Gym-like then it would be easy to make it work with any RL algo. For example, create the env, render the state, take an action as input, etc.

import gym
env = gym.make('CartPole-v0')
env.reset()
for _ in range(1000):
    env.render()
    env.step(env.action_space.sample()) # take a random action
env.close()