Optimizing things in the USSR (2016)

The data problem could have been solved with bar-coding. That's how WalMart got tight control over their inventory. Master Control in Bentonville, Arkansas can see pallet 2345235 leaving warehouse 242 on truck 24343 headed for store 2514. If store 2514 doesn't report arrival of that pallet, questions will be asked and the pallet will be found or somebody will be in big trouble. When the pallet is found, broken down, and product placed on shelves, the items will eventually show up in cash register data. If shelf inventory minus sales is not very close to incoming shipments, the "shrinkage" security people are informed and start looking for theft.

Gosplan had an annual planning cycle and monthly reporting. WalMart has a weekly planning cycle and daily reporting. A year was far too much lag for the system to work well. Although it did for heavy industry, where things take months anyway.

The huge linear programming problem isn't insoluble. It's a sparse matrix; potato chips don't affect steel tubing much. There's probably some way to come close to an optimal solution.[1] Sparse matrix techniques were in their infancy when the USSR was active, though.

[1] http://www.gurobi.com/resources/getting-started/lp-basics

Modern (commercial and open-source) solvers of linear programming (LP) tackle problems of millions of variables in minutes. Most of those 12 billion variables would be cleared out (e.g. fixed to their lower bound, zero) in a pre-processing step.

Problems with variables that are constrained to be integer can take much longer, but they are orders of magnitude more efficient today than a few years ago.

Spufford also wrote The Backroom Boys (https://www.amazon.com/Backroom-Boys-Secret-Return-British/d...) about British tech after WWII, definitely worth a read even if Red Plenty remains unsurpassed.

I'll also add the Wayback Machine for the (now defunct) website created by the author for Red Plenty: https://web.archive.org/web/20160713031430/http://www.redple...

You can't plan innovation. It is hard to plan for knowledge economy in general (material production is increasingly a smaller part of total GDP — something that early Communist theorists failed to predict).

Capitalism is automatic, self-correcting, parallel optimization process. It is asymptotically optimal.

The article makes me think that the failure of the planned economy system would probably not have happened if they had modern computers (and computing tools we take for granted today). The author thinks this would only be possible in about 100 years because he considers that everything needs to be considered in the model. However, one could make the case for doing central planning of a limited set of goods which are interdependent, of course.

AI being all the rage in the news these days, and pretty much in fashion, who knows, perhaps automatic optimization of a country's economy will start to be talked about as a Good Thing To Have?

<start dancing, comrade...>

Simulate perfectly rational economic actors that are only concerned with the next upstream and downstream hops in the market, have them take into account the "taxes" and "credits" of externalities without any of the pesky politics, and distribute in-simulation "purchasing power" to end consumers in whatever way you deem fit. You could even distribute this end-user power to actual consumers, in order to allow them the freedom to input their desires into the system.

Basically, remove the capitalists from the capitalist system; or, just use the computational model of capitalism without actually giving power to the computational nodes.

I have to think that the major objection to this is that the complexity of a socioeconomy also scales with computational capacity. Planning an economy is quite likely always beyond the capacity of any portion of entities within that economy, for the same knowledge problem.

Running up a rigorous approach to proof/disproof of that would be an interesting exercise, of course.

But it is an interesting example of the point that AI has a way of being all things to all people, when rigorous testing of ideas is not applied. For the would-be communists and socialists it is a road to a perfect planned technocracy, regardless of the fact that this looks less than feasible.

(The article includes a link to an article with such name, a good read for an in-depth discussion / comment trail regarding optimization in the USSR / CCCP:)

http://crookedtimber.org/2012/05/30/in-soviet-union-optimiza...

This is actually pretty fascinating. If we reach a level of automation such that factories are completely automated, I wonder if the data that is gathered from all those factories, along with more powerful computers and algorithms in the future, would allow central planning to work better than no planning. Some of the problems mentioned by the author are caused simply by the data being wrong, because of lying factory managers: automated factories would eliminate that problem at least.

At the very least, I hope that in the future there will be a centrally planned economy managed by the government to provide a decent living standard to most citizens, while minimizing environmental impact. More productive or talented citizens would/could choose to opt to earn more wealth and purchase from an open market.

Even if corporations plans their entire operation it is not the same as central planning, because even if a corporation plans its production it cannot plan it sales. It can only estimate because it operates on a common market.

Estimation of sales plans hiring, production etc, however the actual real sales is a feedback loop into that process, where in a central planned system that feedback loop does not exist. Central planned system will just go on whatever the "sales".

Another parallel to the Soviets: Things like good-ol-boy networks and elite school alumni [membership in The Party] are largely what drives promotion within the higher ranks.

But most importantly you don't need to find an optimal solution. Things like gradient descent and hillclimbing will very quickly find locally optimal solutions. That's basically what capitalism is, prices and business demands propagate and sort of do gradient descent (but often very slowly and inefficiently.) Algorithms like scatter search and novelty search can be used to explore the wider search space and find the global optimum. E.g. "what would the locally optimal solution be if we shut down this factory and built a new factory here, and would it be better?"

I found Red Plenty fascinating and I enjoyed the Crooked Timber seminar that Chris Said links to, particularly Shalizi’s post. I'm still wondering: why is the compute-unit "a modern desktop computer?" Is parallel scaling poor for this kind of problem?

If parallel scaling is poor, might better speedups be available with ASICs? Or are there heuristics that run faster and produce usually-close-to-optimal solutions on this type of problem? I'm just running through some common ways to accelerate solutions for compute-intensive tasks that run slowly on a single-socket of general purpose CPUs. For all I know, none of these avenues are promising.

In the real world, things break constantly for various reasons. And there is a trade-off between efficiency and robustness. If you try to make something very very efficient (i.e. optimal), then you also make it very brittle. So for instance, in the communist economies (not that they wouldn't have other problems, they were dictatorships after all), the planners would decide that it is "optimal" to have this one factory producing something, like computer chips. But then something got delayed there, and everybody who expected the thing to be delivered had to wait. These unexpected delays accumulated through this "optimal" solution.

I think main reason why capitalism works is actually because it is terribly inefficient. It duplicates a lot of work, through competition. But the robustness is much higher. And that's why it's a success and has higher overall economic throughput.

You can see this paradox in large corporations, which are, as many people noted here, centrally planned economies. They suffer from the same problem - human planners are trying to be too optimal, the robustness suffers, and the result is often spectacular internal failures.

It's like with spacecrafts. Of course it would be more efficient if the spacecraft had only one engine, one computer, and so on. But what if it breaks? Then the whole mission fails. That's why it needs backup engines, backup computers, and so on.

I think communism might have worked better if it had given more autonomy to managers to make decisions and used output data for the previous year to decide budgets for each factory/company.

[1] https://en.wikipedia.org/wiki/Project_Cybersyn

Other question is do we really need that efficiency? Robots are more efficient than people, so people are already losing their jobs and it'll continue. Now with planned economy there will be much more jobs to lose. May be it's better for our civilization not to be very efficient, at least at current period.

> It is likely that many recommendations for the broader economy would have been ignored as well. For example, if a computer recommended that the price of heating oil should be doubled in the winter, how many politicians would let that happen?

The heating oil required to maintain temperatures in so many cubic feet of air space in the country, given some quantity of heaters at various levels of efficiency, themselves producable by the same giant calculation. The problem is the dimensions/variables are endless and subject to change based on the uncalculable whims of people. Stopping at any one point is arbitrary

I see comments saying that smartphones will provide the needed data to make this idea work well, but that doesn't do anything to answer the question "Will it be better?"

http://www.human-competitive.org/sites/default/files/deb-myb...

The paper describes a real parameter function optimization problem, but the Gosplan was I think more of a combinatorial one. According to the author, the method described there can also be parallelized.

As an example, let's say you want to go and buy shoes. You go to the shoe store and you buy a pair in your number. You don't actually get to choose, the shoes you can buy have already been made for you, you just have to pay for them the planned price you're expected to pay for the pair of shoes. It's also planned when these will wear out and you'll need to go buy a new pair. In fact, all the pairs of shoes you'll need to buy for the next five years have been planned.

You want something else than brown shoes? Fuck you (except if you're high enough the totem pole that you get a choice between brown and black).

You want high-quality leather shoes? Fuck you.

You want to start your own shoe factory? Fuck you, the state owns all the factories.

You want to make your own shoes with leather? Fuck you, all the leather usage is planned centrally by the state using statistics.

You want to raise your own cow on your own land to have your own leather? Fuck you, fuck you and fuck you - you don't get to own land, or have cows. The state owns all the land and all the cows. In fact, the state has already planned all the usage of the leather it will get from killing cows that are yet to even be born.

And shoes are just one good among many. If the state determines only 10% of the population gets to ride a bicycle, it will produce exactly that amount of bicycles. Central planning never speculatively overproduces, or even innovates, because that hinges on thinking you can sway the market this way or that. You think "hey, I bet people will really want to buy wine-red shirts, instead of white striped shirts", so you try it out and you capture part of the market. Central planning can't account for people coming up with ideas in the middle of a 5-year planning cycle.

Central planning is also a form of totalitarianism. Every single totalitarian system in history has always pivoted (if it didn't start as) to a form of serfdom, with all benefits going to the people on top. It doesn't matter if it's a king, or "The Party", C-level management always looks out for themselves first. A state with total power is a really shitty place to live in for most subjects.

That said, you could probably have some sort of rationally chosen economy just for the state's own needs, i.e. the state takes care of the roads, so it needs this much asphalt, so taking a survey of the available factories, it should order from this and that factory for maximum quality at minimum price. Something like an alternative to public bids.

But a fully centrally planned economy only works with robotic workers that never ever consume or buy anything unplanned. You want to know why corruption was so pervasive? Because people want something more than only being allowed to buy thin brown leather shoes, blue jeans and white striped shirts. Centrally planned economy is something like India's caste system on steroids.

If you still think central planning is a good idea, please talk to some people who lived in the USSR, they're about 50-60 years old today and can still tell you good stories. Some of them speak good English. Then think really hard how you'll make sure your proposed central planning system won't devolve to that level of corruption.

"Problems of the Planned Economy"

https://books.google.ca/books?id=NLRVWaNw76MC&pg=PT28&lpg=PT...

GM updated their paint robots to "bid" on new jobs (i.e. cars). They ended up saving 1.5M a year. Not much, but multiple similar things add up.

i.e. when the price of the product accurately reflects the costs, the markets can be more efficient.