What we're about

DataPhilly is a community run group for anyone interested in gaining insights from data. Topics include (but are not limited to) predictive analytics, applied machine learning, big data, data warehousing and data science. We <3 data!

Join our slack to help plan future events! http://bit.ly/DataPhillySlack
Send us a message: http://goo.gl/RvxB6J !
Found a space we can use for future meetups? https://goo.gl/Ru0eth
Found a speaker for an upcoming meetup? https://goo.gl/9DJxq0
Found a sponsor for our events? https://goo.gl/JLVfqh

Upcoming events (1)

Multi-Agent Reinforcement Learning: Dodging Tragedy of the Commons

Event Schedule:
1. Meeting opens at 5:45pm
2. Introductions (Starts at 6pm)
3. Speaker Event
4. Networking Event (30 minutes)

Speaker Event
Topic: Multi-Agent Reinforcement Learning: Dodging Tragedy of the Commons with Simple Mechanisms
Speaker: Quinn Dougherty

Some problems can be described in terms of states, actions, and rewards. A computer program that maximizes rewards in such an environment by selecting actions is called an agent, and the study of these agents is called reinforcement learning. You can select actions with deep learning, leading the research community to advances in playing Go and autonomous vehicles. Naturally, problems and environments arise that are best thought of as a confluence of two or more such agents, the study of which is called multi-agent reinforcement learning. Meanwhile, over in economics, common pool resources are studied as an approximate prisoner's dilemma: if the collective harvests too much, everyone loses, yet if any individual unilaterally implements a sustainable policy others are incentivized not to follow suit. In the literature this is called tragedy of the commons, but economist Elinor Ostrom took an empirical approach [2][3] and found emergent mechanisms all over the world that caused communities to dodge this outcome. You're asking a natural question: do we want to simulate these environments with multi-agent reinforcement learning, simulate a mechanism suggested by Ostrom, and observe if our agents can dodge tragedy of the commons? In this talk, we will discuss my team's journey through this research question and observe the surprisingly easy interface to Ray's RLlib library [4] for training agents to play multi-player games in python. There will be a follow-along repo, if not a notebook.

[1] https://www.lesswrong.com/posts/LBwpubeZSi3ottfjs/aisc5-retrospective-mechanisms-for-avoiding-tragedy-of-the
[2] https://www.amazon.com/Governing-Commons-Evolution-Institutions-Collective/dp/1107569788/ref=sr_1_1?dchild=1&keywords=governing+the+commons&qid=1632844671&sr=8-1
[3] https://en.wikipedia.org/wiki/Elinor_Ostrom#Design_principles_for_Common_Pool_Resource_(CPR)_institution
[4] https://docs.ray.io/en/latest/rllib.html

Speaker Bio:
Quinn Dougherty is a logician at platonic.systems, working on auditing a new decentralized finance project for the Cardano ecosystem and on formal verification. Previously he was a research intern at the Stanford Existential Risks Initiative profiling how the AGI Safety and Alignment research community should prioritize multi-stakeholder and/or multi-agent scenarios. In 2020 he worked on the hospital traffic forecasting app CHIME and did some python and cloud security work for a startup. Quinn is also a coorganizer at Effective Altruism Philadelphia. More information including socials and contact at quinnd.net.

Networking event:
Join us after the talk for a chance to chat with Quinn and network with the other members of the DataPhilly community.

Photos (117)

Find us also at