Dark Forest Theory and Multi-Agent Reinforcement Learning (2023)

Dark Forest Theory and Multi-Agent Reinforcement Learning (2023)(hal.science)

2 points by hamburgererror 146 days ago | 2 comments

elar_verole 146 days ago |

Each agent has 3 available actions. Performing an action costs the agent resources. Currently, resources are allocated arbitrarily, based on the agent’s perception of how much such an action might cost in a real environment.
1. Broadcast own position - Cost: 7 resources
2. Private message own position - Cost: 5 resources

3. Kill other agent if their position is known to the killer - Cost: 10 resourcese

Maybe I am misunderstanding, but OF course if the only options an agent have is to kill someone it knows (and it costs resources) or make itself known, the better option in the end of the process is to never be known.

hamburgererror 146 days ago | |

It's true that a similar analysis with a possibility for cooperation could be interesting to study.