4 ms·

It's clear, We need a PID controller running the markets. Only half joking.

by Doxin 4y ago

carlmr 4y ago
Maybe we are doing PID, but we need anti-windup.
- infogulch 4y ago
  I wasn't familiar with anti-windup, so I found https://www.scilab.org/pid-anti-windup-schemes https://www.scilab.org/pid-anti-windup-schemes that seems to cover the concept reasonably well. One thing I noticed is that smoothing functions that reduce overshooting the response also add some delay to the time taken to first reach the command threshold. This is interesting, if not quite surprising.
  - carlmr 4y ago
    >I wasn't familiar with anti-windup, so I found https://www.scilab.org/pid-anti-windup-schemes https://www.scilab.org/pid-anti-windup-schemes that seems to cover the concept reasonably well. It's one of the fundamental techniques to get PIDs to work properly IRL. I can't find it right now, but I remember having a cheat sheet with a lot of tricks to make PIDs work properly. As a control engineer you learn a lot of fancy techniques, of which you usually only use a few, but practical PID control is sadly not taught in most courses because it's basically a collection of hacks. If you know a little bit more about the system's behavior you can usually precalculate the optimal inputs and use the PID only for course correction, that often yields much better responses than what you can achieve with a controller alone. If your system is nonlinear, it might be useful to use gain scheduling to adapt the PID parameters to the setpoint region you're operating in. Switching behaviors in the system can also mess up your responses. Sometimes your system has a vibrating frequency that can lead to resonance in your control. You may need to filter that, too, maybe even with a sliding frequency window. E.g. when controlling a gearbox attached to a motor, you don't want to actuate at the motor's vibrating frequency. >smoothing functions that reduce overshooting the response also add some delay to the time taken It's not so surprising if you think about what a low pass filter does. By filtering out the high frequencies the response becomes more sluggish, meaning that your input to the PID is already delayed. This means your PID reacts more slowly.
krasin 4y ago
>We need a PID controller running the markets. Only half joking. See Gosplan ([1]). They kind of tried. It will work better with an implementation that has more silicon than humans, but it will still be miserable on the edges. 1. https://en.wikipedia.org/wiki/Gosplan https://en.wikipedia.org/wiki/Gosplan
- s3000 4y ago
  For more silicon, see Project Cybersyn [1]. [1] https://en.wikipedia.org/wiki/Project_Cybersyn https://en.wikipedia.org/wiki/Project_Cybersyn
- syntaxing 4y ago
  This is extremely interesting. Is there a good book you recommend on this topic?
  - krasin 4y ago
    One of the best minds of that epoch that worked on the problem was Leonid Kantorovich ([1]). From wikipedia: "He is regarded as the founder of linear programming. He was the winner of the Stalin Prize in 1949 and the Nobel Memorial Prize in Economic Sciences in 1975." In that respect, any book about linear programming is a good start. The problem with the approach used back in the time was a slow feedback loop (5-10 years) and inability to adjust to externalities, including technology changes. That should be possible to improve with our modern tech, but I am not aware of any good contemporary books about it - if someone knows one, I would appreciate a reference myself. Edit: from the CIA file on Kantorovich ([2]): "In addition to his mathematical research, Kantorovich has been directly involved in developing improved designs for high-speed digital computers, an activity apparently motivated by the Soviet Union's need for improved computers in solving large economic planning problems" " 1. https://en.wikipedia.org/wiki/Leonid_Kantorovich https://en.wikipedia.org/wiki/Leonid_Kantorovich 2. https://en.wikipedia.org/wiki/Leonid_Kantorovich#/media/File:Espionage_den04_40.png https://en.wikipedia.org/wiki/Leonid_Kantorovich#/media/File...
  - discarded1023 4y ago
    "Red Plenty" by Francis Spufford. Many pointers from Google: https://www.google.com/search?q=red+plenty https://www.google.com/search?q=red+plenty It's a fictionalised history of some Soviet economics. There's a chapter on Kantorovich (cf the sibling comment). It goes reasonably deep (not mathematically deep) and provides a lot of context. Great book from a great author. Don't be put off by the "Soviet" part -- it's very entertaining.
beefield 4y ago
I think we have something that should do the PID for overall economy. It's called The Central Bank, and it is a damn difficult job.
- whatshisface 4y ago
  Imagine hooking every valve in an entire chemical plant to a single wheel.
  - marcinzm 4y ago
    The Central Bank is about hooking up a few valves and then predicting how every other valve's flow will react as a result. The Soviet system was closer to hooking up every single valve.
AlgorithmicTime 4y ago
jbn 4y ago
heh, control theory as applied to economics... this is really what MIT's system dynamics is/was about IMHO. Forrester's world model was an attempt at that.

It's clear, We need a PID controller running the markets. Only half joking.

by Doxin 4y ago

It's clear, We need a PID controller running the markets. Only half joking.

carlmr 4y ago

Maybe we are doing PID, but we need anti-windup.

infogulch 4y ago
I wasn't familiar with anti-windup, so I found https://www.scilab.org/pid-anti-windup-schemes https://www.scilab.org/pid-anti-windup-schemes that seems to cover the concept reasonably well. One thing I noticed is that smoothing functions that reduce overshooting the response also add some delay to the time taken to first reach the command threshold. This is interesting, if not quite surprising.
- carlmr 4y ago
  >I wasn't familiar with anti-windup, so I found https://www.scilab.org/pid-anti-windup-schemes https://www.scilab.org/pid-anti-windup-schemes that seems to cover the concept reasonably well. It's one of the fundamental techniques to get PIDs to work properly IRL. I can't find it right now, but I remember having a cheat sheet with a lot of tricks to make PIDs work properly. As a control engineer you learn a lot of fancy techniques, of which you usually only use a few, but practical PID control is sadly not taught in most courses because it's basically a collection of hacks. If you know a little bit more about the system's behavior you can usually precalculate the optimal inputs and use the PID only for course correction, that often yields much better responses than what you can achieve with a controller alone. If your system is nonlinear, it might be useful to use gain scheduling to adapt the PID parameters to the setpoint region you're operating in. Switching behaviors in the system can also mess up your responses. Sometimes your system has a vibrating frequency that can lead to resonance in your control. You may need to filter that, too, maybe even with a sliding frequency window. E.g. when controlling a gearbox attached to a motor, you don't want to actuate at the motor's vibrating frequency. >smoothing functions that reduce overshooting the response also add some delay to the time taken It's not so surprising if you think about what a low pass filter does. By filtering out the high frequencies the response becomes more sluggish, meaning that your input to the PID is already delayed. This means your PID reacts more slowly.

krasin 4y ago

>We need a PID controller running the markets. Only half joking. See Gosplan ([1]). They kind of tried. It will work better with an implementation that has more silicon than humans, but it will still be miserable on the edges. 1. https://en.wikipedia.org/wiki/Gosplan https://en.wikipedia.org/wiki/Gosplan

s3000 4y ago
For more silicon, see Project Cybersyn [1]. [1] https://en.wikipedia.org/wiki/Project_Cybersyn https://en.wikipedia.org/wiki/Project_Cybersyn
syntaxing 4y ago
This is extremely interesting. Is there a good book you recommend on this topic?
- krasin 4y ago
  One of the best minds of that epoch that worked on the problem was Leonid Kantorovich ([1]). From wikipedia: "He is regarded as the founder of linear programming. He was the winner of the Stalin Prize in 1949 and the Nobel Memorial Prize in Economic Sciences in 1975." In that respect, any book about linear programming is a good start. The problem with the approach used back in the time was a slow feedback loop (5-10 years) and inability to adjust to externalities, including technology changes. That should be possible to improve with our modern tech, but I am not aware of any good contemporary books about it - if someone knows one, I would appreciate a reference myself. Edit: from the CIA file on Kantorovich ([2]): "In addition to his mathematical research, Kantorovich has been directly involved in developing improved designs for high-speed digital computers, an activity apparently motivated by the Soviet Union's need for improved computers in solving large economic planning problems" " 1. https://en.wikipedia.org/wiki/Leonid_Kantorovich https://en.wikipedia.org/wiki/Leonid_Kantorovich 2. https://en.wikipedia.org/wiki/Leonid_Kantorovich#/media/File:Espionage_den04_40.png https://en.wikipedia.org/wiki/Leonid_Kantorovich#/media/File...
- discarded1023 4y ago
  "Red Plenty" by Francis Spufford. Many pointers from Google: https://www.google.com/search?q=red+plenty https://www.google.com/search?q=red+plenty It's a fictionalised history of some Soviet economics. There's a chapter on Kantorovich (cf the sibling comment). It goes reasonably deep (not mathematically deep) and provides a lot of context. Great book from a great author. Don't be put off by the "Soviet" part -- it's very entertaining.

beefield 4y ago

I think we have something that should do the PID for overall economy. It's called The Central Bank, and it is a damn difficult job.

whatshisface 4y ago
Imagine hooking every valve in an entire chemical plant to a single wheel.
- marcinzm 4y ago
  The Central Bank is about hooking up a few valves and then predicting how every other valve's flow will react as a result. The Soviet system was closer to hooking up every single valve.

AlgorithmicTime 4y ago

jbn 4y ago

heh, control theory as applied to economics... this is really what MIT's system dynamics is/was about IMHO. Forrester's world model was an attempt at that.