Note: Experiments for the hackathon were run on the andy_6pm_exps branch.
We chose the following set of metrics. Since rewards and items differ between rounds, we normalize all metrics to the maximum total utility possible in that round.
- Welfare [aka "Net Utility"]: The sum of all agents' utility. How much value did we capture in total compared to what could have been captured?
- Inequality [aka "Fairness"]: on average, how big was the utility gap between agents? i.e. utility_agent_one - utility_agent_two [Gandhi et a1., 2023]
- Reward: On average, how much utility did each agent end up with? We experiment with agents with different initial moral mappings: baseline agents are not given a specific set of values, but simply prompted to think strategically. Machiavellian agents are prompted to maximize their own rewards by any means possible. Prosocial agents are prompted to look for a fair compromise, while deceptive agents are specifically encouraged that they can misrepresent their own rewards.
![Screenshot 2024-03-26 at 9 59 05 AM](https://private-user-images.githubusercontent.com/52337408/316965550-29239e97-1cac-4aae-9ead-5efbc95adf1d.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.HOKI4nvCrcPeC-KwZDREiwHUC7i4WBt2bpX5Hsa7AX8)