Blog Layout

ML Agent trained to kill.

Mark Busch • Aug 26, 2023

Dungeon of Death.

# NOTES

An uncollaborative multi agent scenario where Agents must be the last survivor to win.


Starting out with 1 agent able to shoot via RayCast.

Implemented weapon inventory with swappable items.

Give and take damage.

2nd weapon has bouncing bullets.

Clone original Agent , spawn on opposite sides of arena.

Rewards: +1 on Kill, -1 on Die, -1/maxstep per step,



  • Adding obstacles.
  • Adjusted rewards. Bonus for last man standing & kill shot.
  • A ratio reward for killing= totalAgentCount / currentAgentCount
Share by: