Skip to main content

Research and Interests

The fundamental basis of AI and ML is in modeling/representation of the world, and the processing of those representations. The bases of these models are prior knowledge, and data. How does data inform knowledge? How does knowledge inform action? These are questions that go beyond statistics and data science.

Over the last 25 years, an explosion of data and computational resources have pushed AI to impressive advancements in many areas. But ultimately data can only take you so far, and not everything can be understood through correlation and correlation of correlations. So where do we go next?

Interests

  • General Topics
    • Bandit Problems, Optimal Learning Theory
    • Model-Informed Inference
    • Learning on Small to Modest Data Sets
    • Reinforcement Learning
    • Multi-Agent Coordination, Learning, Collaboration
    • Stochastic Optimization
    • Inference Under Uncertainty, Bayesian Inference
    • Unsupervised, Structural Learning
    • Gender in Math and Computer Science Education

Research Lab

There are a couple of problems of active interest.

  • Model-Informed Learning: In AI/ML, learning is often pursued from a model-free perspective, assuming as little as possible about an environment or system. Frequently this is an attempt to eliminate human biases from entering a system. However, there are limitations to what model-free systems can learn and do – for instance, it can have no real notion of causal relationships. To what extent can models be used to inform and direct learning / inference? To what extent can models be learned? What kind of meta-models might be applied?
  • Joint Bandit Problems: In learning, actions are frequently treated as independent from one another: the results of taking an action are generally taken to say anything about the potential results of other actions, like operating two independent slot machines to figure out which one has a higher win rate. But given additional information about how the actions relate to one another, they potentially become informative about each other. If you knew for a fact that one of the slot machines yielded a jackpot 10% of the time and the other only 5% of the time, data on one would suggest the identity of the other. How can joint information be utilized to accelerate learning? This relates to Model-Informed Learning above.
  • Multi-Agent System Fragility: In environments where multiple agents interact with each other and their environment, the actions these agents pursue can have dramatic consequences on themselves, each other, and the environment they are in. We can see this for instance in economics and game theory, with resource management and consumption. In certain situations, individual agents acting in their own interests can destabilize systems (for instance, Tragedy of the Commons) and lead to negative outcomes for all. Where do these instabilities or fragilities come from? How can they be managed or averted? Can these systems be made more stable and beneficial for everyone, and if so, how?
  • System Control via High Performance Actors: We’ve all observed how one bad driver can ruin traffic for everyone else around them. Is the inverse possible? Can one exceptionally good driver, or at least a small team of high precision drivers, improve net traffic over all for everyone? How could these precision drivers be designed and organized for maximum benefit? In multi-agent, decentralized systems, can the actions of few benefit the many?

In general, systems where multiple agents are interacting with each other provide fascinating examples of rich dynamic behavior, and can be very difficult to control. How can we use the tools of AI and reinforcement learning to analyze, control, and benefit these kinds of systems?

As a more speculative project: One of the things that allows humans to be successful, socially, is a theory of mind: a model we use to understand the minds of, and anticipate the actions of, those around us. How could AIs be equipped with a theory of mind for each other, to facilitate action and learning in multi-agent settings?

Publications