Multi-Agent Reinforcement Learning in Zombie-Survival Games

Peter Hollows, Mark Presser · 2020 · Stanford University machine-learning
Download PDF Code

Discussed on the blog: Ultimately the Survivors Do Not Prevail

Abstract

We investigate challenges of agent design using a multi-agent zero-sum game called “Zombie-Survival” in which we train 7 agents (2 survivors, 5 zombies) under different scenarios using the same reward function to investigate how different environmental dynamics can lead to very different learned policies. We show how state-space compression can improve the stability of a learned policy, how a health mechanism that impacts maximum speed can change the dominant strategy, and how arming the survivors with projectile weapons can lead to a social dilemma. Ultimately the survivors do not prevail.

@misc{hollows2020multiage,
  author = {Hollows, Peter and Presser, Mark},
  title  = {{Multi-Agent Reinforcement Learning in Zombie-Survival Games}},
  year   = {2020},
  month  = dec,
  note   = {Stanford University},
  url    = {https://dojo7.com/papers/zombie-marl/}
}