The 'Better Siri' that Could Plan NASA Missions and Help You Organize Your Day

​ Deciding whether to grab lunch at the burrito joint a couple blocks away or the Portuguese sandwich shop across the street can sometimes feel like planning a NASA mission—weather, budget, and time constraints can all play a factor in determining your chances of making it back to work on time with a delicious meal in hand.

Thankfully, risk-assessing artificial intelligence is getting closer to helping both rocket scientists and hungry desk jockeys make the right call in tight situations.

Brian Williams, a researcher at MIT's Computer Science and Artificial Intelligence Laboratory ( CS​AIL) co-invented Remote Agent, an AI NASA used to control​ its Deep Space 1 craft for two days in 1999. Now, he and his team at MIT are looking into eventually putting similar algorithms on your phone as a kind of "better Siri," as Williams put it in a s​tatement.

The algorithms, which his group will present at the annual meeting of the Association for the Advancement of A​rtificial Intelligence (AAAI) this month, work by reasoning within a set of given goals and constraints, offering plans of action above a set threshold of risk. As your mission constraints evolve, the algorithm also changes its recommended plan of action.

"At first, you give it an initial goal and the algorithm just runs in a linear process, but you're also given a chance to review it before it executes the plan," said Andrew Wang, a graduate student in Williams' group. "That gives you a chance to add more constraints, or even remove constraints. As you do that, the system doesn't need to re-solve the problem, it can instead make incremental changes."

This approach to AI problem solving is an improvement over previous methods, Wang told me, because it allows for a certain level of risk and deals in probabilities instead of absolutes. For example, if you tell the algorithm that you want to catch a bus with a 95 percent chance of success, the algorithm assumes that you're fine with a 5 percent chance of having to take a taxi or walk instead, allowing it to reason within those constraints and offer alternative plans if needed.

It's also conversational, in a way, and allows for several stages of reconsideration and compromise as the operator and algorithm work out the best way to achieve their goals. You can imagine telling Siri—to return to this metaphor—that you want to make it to the burrito restaurant in 10 minutes, be back to work after 25 minutes, and stop to get a coffee on the way. The algorithm might then ask if you can extend your time constraint to 35 minutes to fit everything in. If you answer that it's not possible, the algorithm might then suggest you get the Portuguese sandwich instead.

The situation would be largely the same for NASA mission planners using the algorithm, Wang said, although the stakes might be a little bit higher than deciding what to stuff your face with.

"Based on our intuition it seems like these NASA missions are more critical, but really all you have to do is constrain your problem that much more in terms of your probabilistic guarantees, and then the algorithm in principle will work the same [as in everyday situations]," Wang explained.

Wang also told me that his group is interested in implementing the system in cars, not just as a navigational tool, but as a safety measure. "When someone is at risk of falling asleep at the wheel, to have something that keeps your mind focused on the task of driving and getting to their destination could be an attractive safety feature," Wang said.

Though Williams and his team, Wang included, are hard at work on algorithms that can help you plan your daily activities within a reasonable bracket of acceptable risk, there is still much work to be done on the research side. Until then, I guess we'll have to settle for slowly and loudly saying to our phones, "Where. Is. The. Nearest. Burrito."