Suggestions
Andreas Haupt
Economics and Machine Learning. PhD Candidate @MIT. Former @FTC @European Commission @Bundestag @TeachFirstDE.
Social Science of and with (Learning) Algorithms.
Highlights
I'm looking forward to contribute to a panel with Youssef Shaker and Anushree C. for the (small, but hopefully meaningful) delegation of the Massachusetts Institute of Technology to COP28 UAE.
As I look back, I am happy to report on what my visit to COP26 - UN Climate Change Conference 2 years ago led to: A Nicole Immorlica and Brendan Lucier on the design of voluntary carbon markets.
💡 One of the main insights is that there is a conflict of interest between certifiers and market participants: While certifiers are maximizing for objectives such as climate mitigations, market participants maximize gains from trade. This justifies why a laissez-faire policy is not optimal for the climate.
🔎 The impact of ratings, of integration of exchanges with certifiers, and the instrumental effects of co-benefits certification on mitigation activities are direct extensions of our model.
Here is a relatively recent version of the paper, with an update this month: https://lnkd.in/eh_M_WGm.
I'm happy to announce that our team (Dylan Hadfield-Menell Mehul Damani Phillip Christoffersen) at MIT CSAIL has open-sourced our implementation for training multi-agent systems using formal contracts. Access the code repository here: https://lnkd.in/ex-KBvUp.
The code replicates and extends the experiments from "Formal Contracting Mitigates Social Dilemmas in Multi-Agent Reinforcement Learning" (https://lnkd.in/ec2Gsymy). Our algorithms enable agents to make binding commitments through state-action dependent reward transfers, thus achieving socially optimal outcomes without the necessity for explicit reward design.
This codebase is designed to be compatible with any multi-agent domain in Ray's RLLib. We've already included pre-implemented examples in domains such as Harvest, Cleanup, and autonomous vehicular navigation.
To extend the code to new domains, simply modify one of the existing run scripts and define your own contract space logic. For a comprehensive guide, consult the repository's README (https://lnkd.in/etCwXHcb).
We are confident that this open-source contribution will serve as a valuable resource for ongoing research in effective multi-agent systems. We look forward to your active engagement with our work. Your feedback and contributions are highly anticipated.
