Comment Re:Microsoft actually did something good for once (Score 1) 113
MIR (Mode Identification and Reconfiguration) is the part of the Remote Agent that does diagnosis. Here's a shot at explaining how it works. (MIR was designed at NASA's Ames Research Center, although Remote Agent was a joint Ames/JPL project.)
MIR does not use a Bayesian network but comes mostly out of a closely-related thread of AI research called Model-Based Reasoning.
The basic task MIR is solving is, given sensor readings from the spacecraft, what state is each of the components in (e.g., working, degraded, faulty, stuck closed, burned out, and so on). This is tricky because there are relatively few sensors and the state of most components can't be observed directly, but only inferred from the sensors and an "understanding" of how the spacecraft hardware works.
MBR is based on structural and behavioral models of the system, much like a schematic with simulation models for the components. Components are modeled as being on one of several modes, some saying what happens when the component is working (e.g., valve-open, valve-shut) and some, when it is broken (e.g., valve-stuck-closed). In MIR's case, the "simulation model" is based on propositional constraints which allows inputs to be inferred from observed or inferred outputs as well as outputs from inputs. Connecting all the component models together give a large constraint network.
Some "nodes" in the constraint network correspond to sensor readings (e.g., battery-voltage-is-nominal). Given the available sensor readings (which are converted from real voltages to propositional statements like battery-voltage-is-nominal by code that interfaces MIR to the flight control software) MIR then finds component modes that are consistent with those readings. There may be more than one set of component modes that are consistent (e.g., the battery-is-dead or the voltage-sensor-is-broken). Here's where the basian inference comes in. Probabilities are associated with the component modes, and the most likely modes (consistent with the sensors) are picked.
Also, Here's another way to think about the question of "is it AI"? In the early drafts of press releases, we've actually tried to avoid the issue a little by saying "fruit of AI research". The people who did the work identify themselves as AI researchers, go to AI conferences (mostly) and based the work on 20 years of work by lots of people, most of whom called themselves AI researchers. It's called AI as much for its roots as its actual behavior in space.
- Mark Shirley
(I'm a MBR researcher in the Ames group that did remote agent, although I didn't work on RAX directly.)
p.s., Microsoft did indeed do something smart by hiring David Heckerman (and Eric Horvitz), who are excellent researchers. And their work did have an impact on RAX, although an indirect one.
MIR does not use a Bayesian network but comes mostly out of a closely-related thread of AI research called Model-Based Reasoning.
The basic task MIR is solving is, given sensor readings from the spacecraft, what state is each of the components in (e.g., working, degraded, faulty, stuck closed, burned out, and so on). This is tricky because there are relatively few sensors and the state of most components can't be observed directly, but only inferred from the sensors and an "understanding" of how the spacecraft hardware works.
MBR is based on structural and behavioral models of the system, much like a schematic with simulation models for the components. Components are modeled as being on one of several modes, some saying what happens when the component is working (e.g., valve-open, valve-shut) and some, when it is broken (e.g., valve-stuck-closed). In MIR's case, the "simulation model" is based on propositional constraints which allows inputs to be inferred from observed or inferred outputs as well as outputs from inputs. Connecting all the component models together give a large constraint network.
Some "nodes" in the constraint network correspond to sensor readings (e.g., battery-voltage-is-nominal). Given the available sensor readings (which are converted from real voltages to propositional statements like battery-voltage-is-nominal by code that interfaces MIR to the flight control software) MIR then finds component modes that are consistent with those readings. There may be more than one set of component modes that are consistent (e.g., the battery-is-dead or the voltage-sensor-is-broken). Here's where the basian inference comes in. Probabilities are associated with the component modes, and the most likely modes (consistent with the sensors) are picked.
Also, Here's another way to think about the question of "is it AI"? In the early drafts of press releases, we've actually tried to avoid the issue a little by saying "fruit of AI research". The people who did the work identify themselves as AI researchers, go to AI conferences (mostly) and based the work on 20 years of work by lots of people, most of whom called themselves AI researchers. It's called AI as much for its roots as its actual behavior in space.
- Mark Shirley
(I'm a MBR researcher in the Ames group that did remote agent, although I didn't work on RAX directly.)
p.s., Microsoft did indeed do something smart by hiring David Heckerman (and Eric Horvitz), who are excellent researchers. And their work did have an impact on RAX, although an indirect one.
