Dota 2 Creep Block Machine Learning (RL)

Successfully used Reinforcement Learning to train a Continuous Policy Neural Network to do creep blocking.

State consisted of 8 values: the x,y offset of the 4 creeps relative to the hero
Action consisted of 60 values: 20 x 3 parameters (weight, mean1, mean2) which defined 2D normal distributions with fixed variance of 5 and correlation of 0. The x,y offset the hero should move are sampled from these distributions
Reward is given whenever a creep's move distance between t-1 and t is less than a certain value

Link to Thread: http://dev.dota2.com/showthread.php?t...
Link to Code: https://github.com/BeyondGodlikeBot/C...

• #people_blogs
• #dota_2
• #reinforcement_learning
• #machine_learning
• #bot