search

kimbring2 / MOBA_RL

Deep Reinforcement Learning for Multiplayer Online Battle Arena
MIT License
74 stars 9 forks source link

[Suggestion] Zero Sum Rewards #6

Closed Passerby closed 6 months ago

Passerby commented 1 year ago

Good job! Did you complete this project on your own? I believe that converting this into a self-play zero-sum game would yield excellent results. and remove derk.

kimbring2 commented 1 year ago

@Passerby Hello!! Thank for interesting to my project.

I could implement by combining various open source project and train the 1:1 agent.

Theoretically, it is correct to use zero-sum, but the game time is too long and there is no way to get a replay of human expert. Therefore. I had no choice but to give some an additional reward to agent when it hits, kills the enemy.