igorcheb/REINFORCE-LunarLanderContinuous-v2

Reinforce Agent playing LunarLanderContinuous-v2

This is a custom REINFORCE RL agent. Performance has been measured over 900 episodes. To try the agent, user needs to import the ParameterisedPolicy class from the agent_class.py file.
Training progress:

Numbers on X axis are average over 40 episodes, each lasting for about 500 timesteps on average. So in total the agent was trained over about 5e6 timesteps. Learning rate decay schedule: torch.optim.lr_scheduler.StepLR(opt, step_size=4000, gamma=0.7). Training code is shown in the training.py file for reference.

In case video demo does not work, here's a gif:

Minimal code to use the agent:

import gym
from agent_class import ParameterisedPolicy

env_name = 'LunarLanderContinuous-v2'
env = gym.make(env_name)
agent = torch.load('best_reinforce_lunar_lander_cont_model_269.402.pt')
render = True

observation = env.reset()
while True:
    if render:
        env.render()
    action = agent.act(observation)
    observation, reward, done, info = env.step(action)
    
    if done:
        break
env.close()