A short overview of how I utilized deep reinforcement learning to train an agent for playing Battlesnake and reach the top 25 rank (90th percentile).

0:00 Intro
0:21 Rules
1:14 Q-learning
2:10 Game State
2:35 CNN
3:05 Double DQN
3:20 State Augmentation
3:45 Look Ahead
4:01 Results
4:25 Ways to Improve

Links:
play.battlesnake.com/leaderboard/standard-duels/conlan/stats
github.com/conlan/BattleSnakeArena

Articles / Papers:
github.com/ArthurFirmino/gym-battlesnake
medium.com/asymptoticlabs/battlesnake-post-mortem-a5917f9a3428
arxiv.org/abs/1509.06461
arxiv.org/abs/1511.05952

Hyperparameters:
Batch Size: 256
Gamma: 0.999
LR: 0.001
Min Epsilon (training): 0.01
DDQN sync: 10_000 steps
DDQN learn: 3 steps

Music by: Bensound.com/royalty-free-music
License code: W6D2FJKXJ9DCDAFF

Pavlov experiments with dog:
wikipedia.org/wiki/File:Pavlov_experiments_with_dog_Wellcome_M0014738.jpg

#machinelearning #reinforcementlearning @Battlesnake

свернуть