Toy Montezuma's Revenge as Pycolab and Gym environments (working in progress, NOT FULLY USABLE)
APACHE-2.0 License
This is a reproduction of the Toy Montezuma's Revenge environment described in Deep Abstract Q-Networks (Roderick et al., 2017).
Rewards:
Demo (play by hand):
python -m mr_pycolab.toy_montezuma
# All rooms are fully observable rather than partially:
python -m mr_pycolab.toy_montezuma --full-observation
import mr_pycolab, gym
env = gym.make("ToyMontezumaRevenge-v0")
s = env.reset() # [11, 11, 5]
actions = ('D', 'U', 'L', 'R', '?')
s, r, done, info = env.step(env.action_space.sample())