How to make a Steel Producer Angry
Simulated Annealing (SA) is one of the few probabilistic techniques that can search probabilistically for a global solution that just “works” for a complex system. The name “Annealing” comes from “Steel Annealing” which is basically a random method to adjust steel production temperature from heat to cold and so on. Steal Annealing, ironically, isn’t based on a theoretical logic or a specific set of rules but rather it is a meta-heuristic: a trial & error technique that makes things “work”, developed under a long time of experience through acquiring some specific skills to achieve it.
Hence, if we want to propose to a steel maker a certain kind of optimization, due to our smartness and our “zeal” to make things better using odd words like “efficiency”, “optimization” or “stabilize” then chances are high the meeting won’t last for two minutes before being thrown out. The argument here is that a steel maker doesn’t care about Stability under a general optimum, which modernity and OEMs told us about, but rather about the change of States from bad to good in order to fulfill a great product quality at the end of the day. This can only be achieved by the experience of the steel maker, specifically by how many mistakes experienced during his time at steel making.
Toxicity of Optimization & Over-Optimization
Complex systems simply, and sorry to say that, hate Optimization & Stability. Optimization is good for production lines that are not associated with Fat Tails events, rare events on the left side of the probability distribution, such as potato packing machines, juice packaging and so on. Bringing Optimization and over-Optimization to Fukushima power plant has proved to be the main culprit behind the destructive event that happened in 2011; both safety systems failed to prevent the tragedy. Recently, Fukushima undertook some drastic changes from the sizing of the nuclear reactor by simply replacing it by eight smaller reactors: distribution of risks is one of the techniques that minimize overall risks and the achievement of convexity in payoff that we will discuss later in another article. For us, we started squidbot.io to propose alternative algorithms that can provide the variability under a probabilistic domain for the control or safety system that can make plant operation more robust to risky events. We plan to achieve that by simply minimizing overall mistakes rather than just speaking about naive optimization techniques.
Optimization will be Extinct from History like Transfats & Carbs
Optimization, thankfully, is doomed to extinction, not only in heavy manufacturing but practically in any field from medicine to economics and from politics to social studies.
A heavy-industry producer, from a business sense, isn’t better-off by aiming for short-term average (optimization-based) profits but rather, long-term profits that are the results of long-term trial & error.
To go further more into the details, what have we been taught in academe or have been bought into from the software market as in ‘Optimization’ or the worst one, ‘Stability’, is immensely different to what reality in heavy manufacturing shows. And the main idea is that reality contains what is called second-order effects, an aphorism from probability and statistics, that simply explains that reality, because of its higher dimensionality, produces rare events that can be fatalistic in complex heavy manufacturing apart from unexplained downtime, to huge losses in production and possibly to personnel injury. These rare events, sadly, cannot be detected in any optimization whatsoever and I will keep repeating that to the last breath.
Don’t cross a River if it’s 1-meter deep on Average
It is easy for me to tell you that you can cross a 1-meter average deep river as a 1st order information but I still can hide from you the 2nd order effect: somewhere, the river has a sudden 10-meter steep deadly bottom.
To conclude this part, and as we have explained in a previous post, optimization will get you at the mean of things: that is, you just get the average profits, not the sky-is-the-limit ones but more sadly, you don’t get to hide yourself from the rare events. Rare events are simply located on the left of the distribution, that is, far from the mean. To get them, one needs to forget about system identification or brute-force physical simulation (imagine the stupidity of simulation a distillation tower while fighting with operators) and instead, to focus only on understanding what each measurement at any time is conveying as of valuable info. Valuable info can be either of mistakes being done previously or simply, an opportunity to score a higher payoff or production profits, that is, learning the left and right side of the distribution.
Thankfully, squidbot.io is supported enough by assets that can help heavy manufacturers achieve this goal, either from hardware and super-computing capability or from the scientific advisory we are proud to have.
At the end of the day, no one will accept eating the same food everyday, further, as complex organisms, we won’t be glad to trade a hike in the mountain for a boring treadmill. Complexity is much better understood in the variability of conditions, whether bad or good, around it.