So many projects!

by Sebastian Benthall

I’m very happy with my current projects. Things are absolutely clicking. Here is some of what’s up.

I’ve been working alongside DARPA-recipients Dr. Liechty, Zak David, and Dr. Chris Carroll to develop SHARKFin, an open source system for Simulating Heterogeneous Agents with Rational Knowledge and the Financial system. SHARKFin builds on HARK, a system for macroeconomic modeling, but adds integration with complex ABM-based simulations of the financial system. The project straddles the gap between traditional macroeconomic theory and ABM-driven financial research in the style of Blake LeBaron and Richard Bookstaber. I’m so lucky to be part of the project; it’s a fascinating and important set of problems.

Particularly of interest is the challenge of reconciling the Rational Expectations assumption in economics — the idea that agents in a model know the model that they are in and act rationally within it — with the realities of the complexity of the financial system and the intractability that perhaps introduces into the model. The key question we seem to keep asking ourselves is: Is this publishable in an Economics Journal? Being perhaps too contrarian, I wonder: what does it mean for economics-driven public policy if intractable complexity is endogenous the the system? In a perhaps more speculative and ambitious project with Aniket Kesari, we are exploring inefficiencies in the data economy due to the problems with data market microstructure.

Because information asymmetries and bounded rationality increase this complexity, my core NSF research project, which was to develop a framework for heterogeneous agent modeling of the data economy, runs directly into this modeling tractability problems. Luckily for me, I’ve been attending meetings of the Causal Influence Working Group, which is working on many foundational issues in influence diagrams. This exposure has been most useful in helping me think through the design of multi-agent influence diagram models, which is my preferred modeling technique because of how it naturally handles situated information flows.

On the privacy research front, I’m working with Rachel Cummings on integrating Differential Privacy and Contextual Integrity. These two frameworks are like peanut butter and jelly — quite unlike each other, and better together. We’ve gotten a good reception for these ideas at PLSC ’22 and PEPR ’22, and will be presenting a poster about it this week at TPDP ’22. I think influence diagrams are going to help us with this integration as well!

Meanwhile, I have an ongoing project with David Shekman wherein we are surveying the legal and technical foundations for fiduciary duties for computational systems. I’ve come to see this as the right intersection between Contextual Integrity and aligned data policy initiatives and the AI Safety research agenda, specifically AI Alignment problems. While often considered a different disciplinary domain, I see this problem as the flip side of the problems that come up in the general data economy problem. I expect the results, once they start manifesting, to spill over onto each other.

With my co-PIs we are exploring the use of ABMs for software accountability. The key idea here is that computational verification of software accountability requires a model of the system’s dynamic environment — so why not build the model and test the consequences of the software in silico? So far in this project we have used classic ABM models which do not require training agents, but you could see how the problem expands and overlaps with the economic modeling issues raised above. But this project makes use confront quite directly the basic questions of simulations as a method: how can they be calibrated or validated? When and how should they be relied on for consequential policy decisions?

For fun, I have joined the American Society for Cybernetics, which has recently started a new mailing list for “conversations”. It’s hard to overstate how fascinating cybernetics is as kind of mirror phenomenon to contemporary AI, computer science, and economics. Randy Whitaker, who I’m convinced is the world’s leading expert on the work of Humberto Maturana, is single-handedly worth the price of admission to the mailing list, which is the membership fee of ASC. If you have any curiosity about the work of Maturana and Varela and their ‘biology of cognition’ work, this community is happy to discuss its contextual roots. Many members of ASC knew Maturana and Francisco Varela personally, not to mention others like Gregory Bateson and Heinz von Foerster. My curiosity about ‘what happened to cybernetics?’ has been, perhaps, sated — I hope to write a little about what I’ve learned at some point. Folks at ASC, of course, insist that cybernetics will have its come-back any day now. Very helpfully, through my conversations at ASC I’ve managed to convince myself that many of the more subtle philosophical or psychological questions I’ve had can in fact be modeled using modified versions of the Markov Decision Process framework and other rational agent models, and that there are some very juicy results lying in wait there if I could find the time to write them up.

I’m working hard but feel like at last I’m really making progress on some key problems that have motivated me for a long time. Transitioning to work on computational social simulations a few years ago has scratched an itch that was bothering me all through my graduate school training: mere data science, with its shallowly atheoretic and rigidly empirical approach, to me misses the point on so many research problems, where endogenous causal effects, systemic social structure, and sociotechnical organization are the phenomena of interest. Luckily, the computer science and AI communities seem to be opening up interest in just this kind of modeling, and the general science venues have long supported this line of work. So at last I believe I’ve found my research niche. I just need to keep funded so that these projects can come to fruition!

Buried somewhere in this work are ideas for a product or even a company, and I dream sometimes of building something organizational around this work. A delight of open source software as a research method is that technology transfer is relatively easy. We are hopeful that SHARKFin will have some uptake at a government agency, for example. HARK is still in early stages but I think has the potential to evolve into a very powerful framework for modeling multi-agent systems and stochastic dynamic control, an area of AI that is currently overshadowed by Deep Everything but which I think has great potential in many applications.

Things are bright. My only misgiving is that it took my so long to find and embark on these research problems and methods. I’m impatient with myself, as these are all deep fields with plenty of hardworking experts and specialists that have been doing it for much longer than I have. Luckily I have strong and friendly collaborators who seem to think I have something to offer. It is wonderful to be doing such good work.