Addressing COVID19 gaps between data, models and decision brings us back to hierarchical PID control, folks!

I’d like to comment about the way how careful we should be when we use data (even if it’s accurate at the source and at its processing steps!), when we build models to extract some dependencies between elements of the data, and ultimately when we make decisions.

Long time ago (approx. 40 years), when my father took over as head of control engineering department in St. Petersburg electrical engineering institute (LETI) from the previous head, Professor Alexander Vavilov, at their school they were excited by exploring the idea of evolutionary synthesis of control systems. One crucial part of this study was the development of theory of structural synthesis, where models of the system at each level of granularity had to be adequate to the criteria of optimal control. (By the way, graphs were essential in those models)

The basic idea was that depending on the level of granularity (or hierarchy) considered by the modeller, the system can have completely different criteria of correctness and/or optimality, hence certain aspects that are significant at small scale may not be important at a larger scale.
A bit like the criteria of control in the national level is not the same as criteria for control at the municipal level, and not the same at the level of local community, and not the same at the level of family units and individual households.
So, because of these differences and clashes of interest between different levels there is a lot of anxiety and misunderstanding in societies.

So, what the relationship with COVID-19?

Well the relationship is direct.

Let’s take the data on Mortality 2017 from the UK National Statistics:

This data shows that the number of deaths across the country in one year is significant – hundreds of thousands – not far from 1M. The relative number of deaths, that we witness now as a result of COVID-19 even if it will hit 10K-20K would be quite small though.

So, we clearly have different perspectives here, one is national (spatial) that stretches across the whole year (temporal), while the other could be local (e.g. an area of population in London) and taken during these 2-3 weeks of March-April. The relative increase in the number of deaths at the national scale is a small bump on the curve. I.e., integrating the number of deaths, caused by respiratory problems thanks to COVID-19, at the national scale will not give much effect to the game of the totals.

However, if we look spatiotemporally at the small scale we may see a significant rise in terms of differential and even proportional response. So, if we are particularly sensitive to these two aspects, differential and proportional, we may actually decide to react with a powerful action.

What we are facing here is exactly what I started with in my blog. We are facing with the different levels of granularity (or hierarchy, whatever we call it). Consider the coarse granularity. From this point of view our Mother Nature in us may say, well, why bother, the integral response (let’s denote it by letter I) is very small, and we look at time intervals of decades, so there is no need for any great change in decision-making. The problems of environmental nature are much more serious.

But let’s go down to the level of individuals, especially those living in the most affected areas of COVID-19. Again, our Mother Nature in us would tell us, that the rise in deaths due to coronavirus is an alarm, it may trigger a disaster, we may lose the loved ones, lose a job and income. What’s happening here? It’s actually that at the lower granularity level, the criteria for decision-making are based on differential and proportional responses (let’s denote them D and P). So, in mathematical terms at different levels of granularity we apply different coefficients, or what engineers call gains, to these aspects P, I and D, and form our decisions according to those gains or criteria of importance.

So, ultimately it is vital that the data we use, and the models which characterise this data in time and in space, where we calculate partial or full derivatives and integrate in space and time, or proportionalise in space and time, must be adequate to the criteria of significance we apply, and lead to corresponding decision-making at the appropriate level.

No doubt, the nations that are harmoniously hierarchical and fractally uniform, may have less problems in matching criteria of optimality with the P, I and D responses brought be the models from the actual data.

Yet, again we face that PID-control seems to rule the world we live in!

Leave a Reply