Prof. Andrew Allman’s research is in the area of process systems engineering, which entails the development of theory and methods for decision making in the design, operation, and control of chemical and energy systems. We are particularly interested in the use of network theoretic tools enabling more efficient decision making, both for decomposition of more traditional optimization and model predictive control frameworks, as well as for multi-agentic AI frameworks. We also have strong expertise in many-objective optimization, which we use to help design and understand the training objectives of ML models and AI systems.
Please describe one or two of your most interesting projects.
Dimensionality reduction for many-objective optimization problems: Optimization problems with more than 4 objectives are commonly encountered when making decisions for sustainability an Despite this, works solving these problems rigorously are rare owing to challenges both in solving the problem (methods scale exponentially with number of objectives) and interpreting solutions (humans struggle to visualize trends in higher than 3 dimensions). Our team has developed methods for predicting, a priori to solving the problem, the likelihood that different objectives are correlating (pointing towards the same solution) or competing (giving a wide range of tradeoffs). We can then use network theoretic tools to group objectives into 2 or 3 groups on the basis of this correlating of competing nature, allowing use to more easily identify and interpret the tradeoff space. We have applied these methods extensively to benchmark many-objective problems and problems in the sustainable process engineering space, and are interested on extending these methods for the development of “aligned by design” AI systems in the near future.
Adaptive learning of decomposition for optimization, control, and multi-agentic AI: Decomposition, whereby a challenging problem is separated into a series of easier-to-solve subproblems which can then be coordinated , is a powerful paradigm for complex decision making problems. The best decomposition of a problem is dependent both on the problem structure (connectivity between model equations and variables), which can be identified and exploited using network theoretic tools, as well as by time-varying problem parameters. We are working on developing graph-based AI models that learn the dependence on the best decomposition as a function of both structure and parameters, enabling an adaptive approach whereby the decomposition used to solve the problem changes in response to changes in system states, set points, and disturbances. We are also developing generative approaches that can propose new, out of distribution decompositions that can outperform those in the current portfolio of options. We are also interested in extending these tools to systems where decisions are made by AI agents, developing self-adapting multi-agentic AI systems which can solve more complex tasks than can be addressed monolithically.
How did you end up where you are today? (Your research journey)
I am one of the rare academics who did not do any research as an undergrad – I applied to grad school mainly because I had no desire to work in industry – I simply wanted to keep learning, be challenged, and not sell out to “the man”. At the time, my research interests were very broadly related to problems in sustainable energy. After joining the department of Chemical Engineering and Materials Science at the University of Minnesota, I very much enjoyed the first semester graduate math course, and got along very well with the faculty teaching the course. When advisor matching occurred, I put him as my second choice of faculty advisor. In hindsight, getting matched to my second choice advisor was probably the best thing to happen for my future academic career. Since then, I have been immersed in research entailing optimization, control, network theory, and their applications to sustainable chemical and energy production systems. The rest, as they say, is history.