Personalized Optimization with Human Feedback – Human-O

Making optimisation work for humans

Often in cyber-physical systems, like smart grids and transportation networks, humans are an after thought. We want to design online optimisation algorithms that consider humans from the start. This is hard to do mathematically, but it will open the way for transparent and flexible human-systems.

Data-driven and flexible optimization algorithms

In this project, I propose to explore and design new algorithms at the intersection of optimiza- tion, dynamical systems, and learning of the user’s costs and constraints, allowing end users to have a choice between possible decisions, and learn from it. [S1] Learning to optimize for dynamical systems. We will study how to go beyond the state of the art by learning and predicting the optimizer trajectory from data. Also, we will see how to incorporate prior knowledge from historical data. Complementary to the dynamical system, we will extend the learning to the cost function and the constraints of the problem, both to better predict how they change in time. [S2] Learning to personalize. We will use machine learning (kernel) methods to learn the human satisfaction, and therefore their costs and (possibly) constraints via their feedback, in a setting that could resemble and go beyond a bandit framework. The first key challenge that we will tackle is the combination of the needs of optimization and the properties of traditional machine learning techniques. In particular, we will study how to impose functional/shape constraints for the estimated cost and constraints. [S3] Learning to be flexible. We will go beyond learning a cost or constraints, and model the problem in a way that the result of the optimization is a set of possible solutions and we will give to the users the choice to pick the solution they prefer. This brings in novel perspectives on how to model flexible and time-varying optimization problems, and how to learn from the choice the users make. In particular, for this objective, humans play a more active role, and I expect their preferences (in terms of choices) to give rise to non-convexities and non-idealities.