From linear programming in 1958 to winning at Go

Neil McNaughton traces the history of AI in oil and gas back to the dawn of the computer age then wonders what DeepMind’s success at Go means for complex problems such as seismic interpretation? Can we learn from AI accidents like Tay or the Google car fender bender? If you can't improve what you can't measure, can you apply AI to a problem without knowing what ’success’ is?

It is a pity that artificial intelligence (AI) guru Marvin Minsky died before witnessing Google’s DeepMind beating the human champ, Lee Se-dol, at the game of Go. On the other hand, he was probably turning in his grave as Microsoft’s Tay became a ‘Hitler-loving sexbot’ within minutes of entering the twittersphere.

DeepMind’s 4 to 1 thrashing of Se-dol came at a great time for the AI marketing brigade and journos the world over, as confirmation of the imminent job destructive power of AI. In oil and gas I have heard people who should know better claim that all we need to do now is dump our data, higgledy-piggledy, in a massive container and let the machine do the rest. It has been suggested that units of measure are no longer worth recording as the machine will figure out such trivia.

To trace AI’s history in oil and gas, I checked back through OnePetro to see how long it has been in use. It turns out that the first mention of AI goes back to 1973 and there have been some 1,500 papers that contain the term since then. Frequency of use of ‘AI’ has increased steadily with an approximate doubling (to 100 papers per year) in the last decade.

A search for ‘machine learning’ returns just over 600 results but pushes back the dawn of AI to 1969. If you allow that ‘linear programming’ is a piece of AI, this brings up 624 results going even further back, to 1958. I therefore nominate and simultaneously award Oil IT Journal’s AI starter prize to Messrs. Lee and Aronofsky of the Magnolia Petroleum Co. for their 1958 paper, ‘A linear programming model for scheduling crude oil production.’

But to get back to the present, it is clear that there is good AI and bad AI. How do you make sure that you (or the folks that are working for you) are developing the next DeepMind and not about to release another Tay?

Winning at Go is different from writing a chatbot for the simple reason that the outcome is clear; a win. This cannot be contested* and is unlikely to upset anybody except for Mr. Se-dol. But a chatbot? How do you define its ‘success?’ To paraphrase the old data dictum, if you don’t know what success looks like you can’t optimize your process.

Roadside cameras that capture your car number plate as you drive along are pretty good now and can be considered a modest success for AI. But what about Google’s other big AI adventure, the Google car? While this is a real gee-whiz use of AI, the recent fender bender is instructive. It seems like the car pulled out slowly into a gap in the traffic ‘expecting’ the bus to give way. It didn’t.

So what will Google’s programmers do next? They could change the program to allow all traffic to pass until there is a big enough space to pull out, which could take a while and drive the passengers crazy. They could make the indicator lights flash more brightly. Unfortunately, my experience of driving in Houston suggests that flashing indicator lights are taken by following traffic as a sign to accelerate and block your maneuver. A similar obstacle to ‘do no evil’ Google is what to do when the lights turn red. Accelerate like everyone else? Or brake legally and have the rush hour traffic tail-end your Noddy car?

Driving is imponderable with regards to an outcome. You can set a simple goal for driving like ‘No fender benders.’ But driving around at 50 kph, giving way to busses would not be much of a satisfactory outcome for the F1 driver who recently expressed a preference for the older 1000hp motors over today’s wimpy 700hp V8s. Maybe Google should be looking to buy a racing car manufacturer, add in its AI and see if they could win a F1 race which would at least be a definitive outcome. We might even see Amazon race Google and Facebook. How about a driverless race of 5000HP super-duper cars with Arria doing the commentary? Would anyone listen?

AI in oil and gas is definitely on the tricky side of the equation. On the one hand, ‘optimization’ is a motherhood and apple pie concept. I mean you are not doing to deliberately aim to do un-optimization. But what are you to optimize?

Say we had complete control over every facet of an oilfield. What do we aim for? Ultimate recovery? Cash flow? NPV? ROI? A job for life? Deciding on your goal is just the start. Next you need to think about what the best course of action is most likely to bring ‘success.’ This is where you start down a road with multiple bifurcations. Choices might include a hypothetical new well, extra compressors, pipelines and so on. Before the real world ‘run’ is through, the oil price will probably have changed. And maybe so will the tax take. Optimizing for recovery may mean cutting back on production for months or years which may be hard to explain to your bankers!

Another seemingly elusive target for AI is evidenced in this month’s lead where we report on an ML-based technique for identifying faults in seismic data. This uses a synthetic training data set to come up with a strategy for fault detection. But is it really getting more ‘successful’ at fault finding? How does the algorithm learn in the face of a nebulous outcome, where expert geoscientists may engage in heated arguments as to the origin of the seismic response? Go it ain’t.

* In fairness to Mr. Se-dol it has been contested. On the basis that sharing the time allowed to either party equally is unfair. Human thought proceeds at the speed of thought, but no faster. You can perform more calculations in a given time slot just by using a bigger computer, which doesn’t really mean that it is getting more intelligent.

The true nature of DeepMind in the context of Minsky’s AI is also debated.


Click here to comment on this article

Click here to view this article in context on a desktop

© Oil IT Journal - all rights reserved.