a collection of environments for the OpenAI Gym

current environments:

• Blackjack-v0

all the environments are registered with the gym when you import gymn_nastics

This is not in pip/pypi yet. Sorry. The easiest way will be to add a subdirectory into your project called gym_nastics by downloading the zip here and extracting it or adding a git submodule.

• currently a single, freshly shuffled 52 card deck is used each episode/game
• supports actions Hit (0) and Stand (1)
• observation is Tuple(Discrete(11),Discrete(11))
• you only get a reward when done==True
• your reward will be the score of your hand if you beat the dealer and didn't bust
• the players opening hand will never be a blackjack, those hands require no decisions
• Splitting and Betting are not supported (yet?)

in some trial runs a version of cem.py scored a 100 episode mean of 10.05 (after episode #639,900)

from __future__ import print_function
import gym
import gym_nastics
# environments are registered upon import

# Support python 2 and 3
try: range = xrange
except NameError: pass

env = gym.make('Blackjack-v0')
for i_episode in range(20):
    observation = env.reset()
    env.render()
    for t in range(100):
        action = env.action_space.sample()
        observation, reward, done, info = env.step(action)
        env.render()
        if done:
            print("Episode finished after {} timesteps\n".format(t + 1))
            break