Dechter, Ihler Work to Advance Probabilistic Reasoning
UCI Computer Science Professors Rina Dechter and Alexander Ihler are collaborating with Charles River Analytics (CRA), which has a contract with the U.S. Air Force to develop probabilistic reasoning tools for satellites. CRA’s work is part of its Probabilistic Reasoning for Enhanced Course of Action Generation (PRECOG) grant, which is focused on research to help satellites autonomously determine the best course of action.
Dechter and Ihler’s area of expertise is domain-independent algorithms for probabilistic reasoning. According to Ihler, domain experts will determine the key variables, and CRA will then craft a model that captures all of those variables. “Our part,” he explains, “is the engine underneath, which, given this model, tries to reason about the current observations to determine what optimal action to take.” As Dechter puts it, “CRA is in charge of developing the software application, and we’re the experts in algorithms.”
The strength of their algorithms is the ability to determine an optimal action in a relatively short time. This is essential for performing real-time analytics on stripped-down platforms with limited computing capabilities. Dechter and Ihler worked with CRA previously on a five-year DARPA grant related to machine learning, and, in that case, their algorithms helped the CRA system reason more rapidly about various problems. “Using our algorithms, it ran much, much faster,” says Ihler.
Furthermore, Dechter says that working on such projects helps advance their research. “It’s very useful for us to have some guidance from real applications.” In theory, their algorithms should be applicable to any domain that fits within their general framework. In practice, however, delivering algorithms that are domain-independent is challenging, because the algorithms might not scale well.
By collaborating with CRA, “we have an opportunity to work on this challenge in the context of a concrete application,” says Dechter. This work will help them determine the scalability of their algorithms in terms of looking farther into the future or considering a greater number of threats or potential actions. Ihler explains, “If they’re okay with three threat levels, then that’s great. If they need 10,000 levels, then that’s going to cause a problem for our algorithm. So having that grounding in reality is very helpful.” It will ultimately help them identify acceptable approximation algorithms as well as those that need more fine-tuning for greater scalability. And all while helping out the U.S. Air Force, thanks to CRA’s 27-month contract.