Organizer(s):

Tamra Carpenter, DIMACS

David Pennock, DIMACS

Segev Wasserkrug, IBM Research

Description

Many real-world decision-making situations involve the decisions of multiple parties. Typically, in such situations each party wants to optimize its own objectives, even knowing that their actions affect the objectives of others and vice versa. Game theory is the mathematical science intended to model such situations. Yet many real decision-making scenarios are far too complex for traditional game theory to provide prescriptive recommendations; for example:

• Business networks: settings where multiple, possibly independent, enterprises cooperate in order to obtain more value as a network than they can as individual firms. An example of such a network is a retail supply chain, consisting of supermarket chains, distributors, farmers, etc. Such settings result in an interesting combination of cooperation and competition: firms cooperate to increase the network value, while simultaneously competing on the division of this value.
• Cloud and hybrid cloud computer systems: Analyzing games in which the players include, but are not limited to: cloud providers, enterprise infrastructure teams, companies using the IT infrastructure to provide services to their customers, and the end customers themselves. Problems in this space include contracting, pricing decisions, infrastructure usage decisions, and workload placement.
• Cybersecurity: optimizing decisions for improving cybersecurity by analyzing such settings as adversarial games between defenders and attackers.

Traditional game-theoretic assumptions, including unbounded rationality, unlimited computation, common knowledge, and common priors, often don’t apply in these settings. Algorithmic game theory relaxes some of these assumptions, examining the behavior of parties and mechanisms with bounded computational resources, while simultaneously trying to provide clear and simple mechanisms for the participants to follow. Still, multiple solution concepts with multiple equilibria dilute the predictions that (algorithmic) game theory can produce, and there are many additional real-world scenarios in which computationally efficient and clear mechanisms have yet to be created. Finally, multiagent reinforcement learning (MARL) and other machine learning techniques offer a more prescriptive approach to teach agents how to behave in multi-party settings. These techniques sometimes work quite well in practice, and in vastly more complex settings than traditional game theory can handle, yet less is known about their theoretical convergence properties and performance guarantees.

It is therefore the goal of this workshop to bring together researchers from both industry and academia in the domains of game theory, algorithmic game theory, multi agent reinforcement learning, and learning in game theory to understand and study the problems in which multi-party decisions are required, create joint awareness of the current state-of-the-art in both industry and academia, including relevant software tools and platforms, and seed collaborations both in the integration and scientific advancement of these techniques, as well as their application to real world use cases such as the ones described above.
Topics of interest include, but are not limited to:

• Real-world business scenarios in which such techniques are required, as well as the theoretical and practical gaps inhibiting their application.
• Solution concepts relevant to such real-world use cases and their efficient computation.
• Building upon commonly used knowledge frameworks such as common knowledge priors and knowledge hierarchies to define knowledge frameworks appropriate for real-world decision-making settings.
• Algorithms enabling agents to learn how to best behave in such settings using algorithms such as MARL and regret minimization, based on sound scientific principles.
• Research on how game theory, algorithmic game theory and MARL algorithms can be successfully applied to scenarios such as cloud computing, supply chain networks and security.

View video playlist.

We are grateful to IBM Research for its generous support of this event.

Program

Thursday, October 27, 2022

Friday, October 28, 2022

Call For Participation

This will be a two-day workshop.

• Day 1: Lectures and discussion on current relevant state of the art both in academia and in industry
• Day 2: Brainstorming and breakout sessions for discussions and seeding collaborations.

Presentations at the workshop are by invitation and will occur on only the first day of the workshop. Attendance at the workshop is also by invitation, but those who would like to attend may request an invitation from the organizers. The Day 2 breakouts will occur in a smaller space, so we will have greater ability to accommodate requests to attend the Day 1 lectures than to attend both days.

Lectures will be recorded and posted if the presenter grants permission to do so.

Parking Information

Visitors may park in Lots 74A, 76, and 82. They must use the link below to register for their event. Until this process is completed their vehicles are not registered and they may receive a citation. Special event parking and special event permits are only for visitors to the University, which does not include free metered parking. Faculty, Staff, and Students must park only in lots they are authorized to park in and should not register using this link.  

https://rudots.nupark.com/v2/portal/eventregister/8d582efb-5be7-40f9-a744-c9acce743e0e#/events/registration/    

Directions to parking lots can be found at http://maps.rutgers.edu

Sponsored by IBM Research, in association with the Special Focus on Bridging Continuous and Discrete Optimization , DATA-INSPIRE TRIPODS Institute and the SF on Mechanisms & Algorithms to Augment Human Decision Making.