Heavy Industry: Payoff is what counts, not Prediction

How not to predict Earthquakes – Optimization, the mother of all Collapses – Wearing a thick coat on a Greek beach in August – Postdicting

Un-predicting Earthquakes

There’s a reason why we never and will never predict earthquakes and that’s due to a very simple idea: asymmetry. We know that it suffices only one bad prediction for our super-predictive earthquake model to have and then we might not be writing this post right now. Earthquakes aren’t classified under a Gaussian world, similarly to your Instagram junk food preferences feed. Under complexity, harm can be enormous and this is something central and innate for the collective. We learned that we can’t interfere in complex systems that can cause Gods anger since antiquity.

Prediction for simple domains, Payoff for complex ones

Enter Modernism. Our life is built today on prediction of preferences, from food to relationships to career choice. The most intriguing one is Weather prediction. I used to check my Weather app everyday couple of years back so I can have an idea if I should opt for the umbrella, the swimsuit or the very thick water-resistant coat. Most of the time, the app was predicting well but from time to time I ended up on the beach with an umbrella or, in some other cases, with a T-shirt under snow. But a mistake from the weather app can affect the collective and not the individual, unlike your dumb Instagram junk food suggestion predictor. The alternative is to simply stick to what our ancestors did: leave home everyday with an umbrella, a swimsuit and a thick coat; in this way, you’re fully covered and you don’t care about any predictive or Big Data analytics. Insurance companies do the same job. So let’s rephrase our central idea:

When domain Scalability increases under complexity, correlations & prediction become meaningless. The focus should be done on Payoff.

Asymmetrical Payoff

So what do we mean by Payoff instead of Prediction and what does it mean for Engineers in Heavy Industry? Simply put, in our first example above, the collective would be better of in spending money on preventing deaths from earthquakes rather than spending on forecasting tools. That is, enforcing citizens to build homes under a specific robust standards or any policy that can help decrease the toll of casualties. In other words, the collective understood the fact that there can’t be nothing done against a complex system but our payoff from it, the number of casualties, can be managed.

Hence the idea is to explore our environment data, and how we interacted with it, and then we attempt to decrease bad decisions through probabilities so our long-term payoff has more upside than downside. This is done through Reinforcement Learning, the future of AI, and some probability techniques based on Markov spaces, but this is for another post.

As an example, an oil producer who wishes to increase production through software, instead of predicting the next production output to adjust his actions based on that future value, he can simply watch the past actual values of that production throughput amount, look for the weak ones and then, using probability to minimize the actions that led to bad results for the long run. This is post-dicting, the anti-thesis of predicting. To understand the idea of asymmetry, that is, choosing between predicting under linearity or increasing one’s payoff under complexity, Nassim Nicolas Taleb puts it clearly:

“For instance, for a linear payoff, if you predict an event priced at 1% probability, you can be wrong 99% of the time. For the same probability w/ a convex payoff, you can be wrong 999 times out of 1000 and still do very well”.

Better Decisions, better long-term Payoff

But there are challenges indeed for us where we need to promote and present our ideas to Heavy Industry personnel in a clear way. To sum it in simple terms, we are in fact doing nothing special nor creating any new technology. We are simply helping to locate weak decisions with high probability in a complex industrial domain and attempting to minimize them on the long-run. Less bad luck is more profits on the long-run; this can be true practically in any field or domain.

At squidbot.io, we are equipped with enough super-computing power that lets engineers analyze their complex industrial data, through the use of the latest techniques in probability, mathematics and reinforcement learning. We strive of course to provide a better understanding onto how engineers can drive & control better their processes, without interference in physical assets such as simulation or system identification but under the fact that complexity shouldn’t be squeezed into a linear model that deprives us from future profits or more importantly, hides from imminent risks. Therefore, super-computing power will be used to analyze each peace of data, bad or good, and use that as an experience for the future.

Fighting your plant operator

I have been into heavy industry for the last ten years, promoting products that are “Efficient”, however, I used to end up fighting with engineers & production operators at the end of the day, during the commissioning of these “Efficient” products. No operator would like the idea of messing with their production data, during a system identification procedure, even if this would be beneficial. But somehow, as I discovered, engineers and operators struggling their day, everyday, know something about Asymmetry: Earthquakes can’t be predicted.