Initial public release (PyCon UA 2017)

Two DQN models pre-trained on 60,000 episodes.

Performance (fruits eaten):

• 10x10-blank: mean 19.2, stddev 4.6
• 10x10-obstacles: mean 15.3, stddev 3.4