http://www.chrisman.org/Lonnie   
 

Research

How can we leverage computers for decision making in situations with little or no historical precedent? This is where you find yourself when considering a fundamentally new project or initiative that has never been attempted, or proposing a new government policy, or stategic planning in general. It may also be that the world has changed, perhaps due to technological advances, changes to the competitive landscape, new regulations, or the onset of a pandemic. In these cases, the past is not predictive of the future, which violates the basic premise behind Machine Learning (my own field of study). For the past few decades, this has been the focus of my own research, and has led to the development of the Analytica, visual modeling software with the unique forte of helping you gain insight with model-based decision in novel decision situations.

In graduate school at CMU (1987-1996), my research focused primarily on aspects of robotic intelligence. My main focus was on acting, planning and learning when an agent has only incomplete and noisy perception of its environment. This include planning of sensing operations, which would often take as much time as physical movement. It was this focus that led me into probabilistic inference, which has has been the unifying component of my career.

Some notable research during that period: One of the earliest uses of POMDPs for probabilistic planning in the AI community; reinforcement learning with incomplete perception; recurrent (RAAM-based) neural network learning; inference using upper and lower probabilities; iterative discretization in Bayesian networks with continuous variables; hybrid MCMC / propagation algorithm for Bayesian networks, especially relevant when there are continuous variables.

After graduate school, at a non-profit organization closely affiliated with Lumina (1997), I developed web-based computational tools for eliciting uncertainty assessments from expert using interactive protocols oriented towards overcoming standard cognitive biases, such as anchoring and overconfidence. The application I developed won an award at a health informatics conference for the most innovative user interface.

Also during the early days of Lumina, a focus of my research and development was on developing web-based advisor applications that would bring the techniques and methodologies of decision analysis, which we used at Lumina to help large organizations make high-stake (e.g., multi-million or billion dollar) decisions down to a level that could help individuals make every-day decisions, such as product selection decisions. This thread led to the Advisor technology, which was acquired by Ask Jeeves in 1999, and also led to my three-year adventure with what was then the start-up internet company ask.com.

At Ask Jeeves, as part of Jeeves Corporate Solutions, I led development of the Advisor technology, but also led the development of Answers from Databases, a product that could locate product information in relational databases from English language questions. Examples were: "Do you have any color laser printers that can hold 500 sheets of paper?". By leveraging the very narrow domain, it actually worked quite well and was deployed on several major and well-known eCommerce vendor websites.

In 2002-2003, following a keen interest in bioinformatics and computational biology, I returned to academic research with Pat Langley's Computational Learning Lab at the Institute for the Study of Learning and Expertise (ISLE), then located at the Center for the Study of Language and Expertise (CLSI) at Stanford University. My research there focused on the combination of background knowledge with experimental data for the purpose of Causal Discovery, particularly the discovery of genetic regulatory pathways within cells. My research there focused on Bayesian causal discovery methods, using Bayesian Network style models of cellular regulation and Metropolis Hastings algorithms for evaluating the weight of evidence for causal hypotheses.