"Feedback control is a means of getting a system to achieve and maintain a desired state, usually called the set point, by continuously comparing its current state with its desired state." (Matarić, 2007, p. 121)
The most popular closed-loop control mechanisms are some combination of proportional, integral, and derivative feedback, together named a PID controller.
These mechanisms all need a measurement of the system’s current state compared to the desired state, called the error. A proportional mechanism sends feedback in direct proportion to the error, mainly influencing the speed of response. Derivative mechanisms use the rate of change of the error (the derivative!), mainly influencing the overshoot, related to system damping. An integral mechanism keeps track of the total sum (the integral!) of the error, which helps get rid of the absolute error.
Various PID controller configurations, used to heat a bundle of resistors, were compared in the blog post PID Without a Clue.
PID controllers are fantastic. You don’t need to model your system, and their use is so widespread that tuning methods have been developed for all sorts of applications, making them an easy choice. They work especially well for single-input, single-output (SISO) linear systems without resonances, and that do not change their behaviour over time (for example, after warming up). It is a great observer mechanism – one with no information about the system itself.
For those multiple-input, multiple-output (MIMO) non-linear systems, and those whose performance need improving, it may be a good idea to look into other feedback mechanisms.
Information overload! We don’t have time to finish our PhD’s in control theory – we have robots to build! What are a few things we can do to improve over a simple PID model?
The first thing one can do is to better tune your controller. Tuning really means setting the coefficients of the P, I, and D terms, setting limits to the derivative and integral terms, and sometimes tweaking your sample rate and drive term update rate. Proven methods include Zeigler-Nichols, Tyreus Luyben, Cohen-Coon, Åström-Hägglund, and using software tools.
The best thing to do is to address the main limitation of PID – it has no knowledge of the process under control – by combining it with an Open-loop or feed-forward controller. These use a system model, often simplified, to predict the appropriate response to an input. Modeling the system allows more intelligent sensor measurement conditioning, especially useful for reducing noise in the derivative term and limiting the effect of the integral term.
There are many other improvements one can make to a PID controller, the above few being relatively easy and having a large impact on performance.
The Landing is a social site for Athabasca University staff, students and invited guests. It is a space where they can share, communicate and connect with anyone or everyone.
Unless you are logged in, you will only be able to see the fraction of posts on the site that have been made public. Right now you are not logged in.
If you have an Athabasca University login ID, use your standard username and password to access this site.
We welcome comments on public posts from members of the public. Please note, however, that all comments made on public posts must be moderated by their owners before they become visible on the site. The owner of the post (and no one else) has to do that.
If you want the full range of features and you have a login ID, log in using the links at the top of the page or at https://landing.athabascau.ca/login (logins are secure and encrypted)
Posts made here are the responsibility of their owners and may not reflect the views of Athabasca University.
Comments
Thanks Tyler. Your article stimulated an interest in intelligence control for me.