University of Cambridge > Talks.cam > Engineering Safe AI > Dynamic Safe Interruptibility for Decentralized Multi-Agent Reinforcement Learning

Dynamic Safe Interruptibility for Decentralized Multi-Agent Reinforcement Learning

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Adrià Garriga Alonso.

One potential approach to the value alignment problem is to build for corrigibility: to try to construct a system for which we can modify its operation at any point. This should be the case even if its objectives are incorrect or would be harmed by the modification.

To this end, this week we read “Dynamic Safe Interruptibility for Decentralized Multi-Agent Reinforcement Learning”, by El Mahdi El Mhamdi, Guerraoui, Hendrikx and Maurer. They extend the notion of interruptibility to multi-agent algorithms: they construct a way of conditioning a multi-agent reinforcement learner such that the agents won’t learn to “plan around” interruptions of its operations. In essence, they will act as if they believed they would never be interrupted. This is based on the initial, single-agent, case by Armstrong and Orseau [2], which we won’t read this week.

There will be free pizza. At 17:00, we will start reading the paper, mostly individually. At 17:30, the discussion leader will start going through the paper, making sure everyone understands, and encouraging discussion about its contents and implications.

Even if you think you cannot contribute to the conversation, you should give it a try. Last year we had several people from non-computer-y backgrounds, and others who hadn’t thought about alignment before, that ended up being essential. If you have already read the paper in your own time you can come in time for the discussion.

A basic understanding of machine learning is helpful, but detailed knowledge of the latest techniques is not required. Each session will have a brief recap of immediate necessary knowledge. The goal of this series is to get people to know more about the existing work in AI research, and eventually contribute to the field.

Invite your friends to join the mailing list (https://lists.cam.ac.uk/mailman/listinfo/eng-safe-ai), the Facebook group (https://www.facebook.com/groups/1070763633063871) or the talks.cam page (https://talks.cam.ac.uk/show/index/80932). Details about the next meeting, the week’s topic and other events will be advertised in these places.

[1] (to read) El Mahdi El Mhamdi, Guerraoui, Hendrikx and Maurer. “Dynamic Safe Interruptibility for Decentralized Multi-Agent Reinforcement Learning”. https://arxiv.org/abs/1704.02882

[2] “Safely Interruptible Agents”, Stuart Armstrong and Laurent Orseau. http://intelligence.org/files/Interruptibility.pdf

This talk is part of the Engineering Safe AI series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.

 

© 2006-2018 Talks.cam, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity