Evolved Fuzzy Controllers

Autonomous training method for Fuzzy Logic Controllers using Genetic Algorithms in a Multi-Agent Environment. The parameters are started at random and the selection criteria is the number of captured greens. Full paper here: http://www.lsi.usp.br/~rponeves/

Category:

Science & Technology

Tags:

• Genetic Algorithm
• Evolvable Hardware
• Control Systems
• Fuzzy Logic

[sabriath]

Maybe I'm wrong, but last I checked "Fuzzy Logic" means a perception based on personal outlook. For example, I'm 5'8", a man who is 6'0" would be "tall" to me, and considering the average male is about my height, 6'0" would be "tall" in consideration even to a 6'8" person. However, 5'9" person begins to get "fuzzy" in tallness category. Where does the GA in this film about eating show fuzzy logic? A rename to "GA evolution" would probably be better.

[d4sp1d3r]

You are not wrong, but there is more to the theory than only that, including practical applications, such as Fuzzy controllers.

In the experiment, GA is used to find the optimal crispy boundaries for several fuzzy groups. The selected parameters are used in the decision making circuitry.

[sabriath]

So basically you are saying that they see their world as "the food is near you" or "the food is far to the right of you" and they have to figure out what that means? If this is true, you are taking a probabilistic encoder to make a bunch of agents learn how to decode it, adding an extra step in the process, now it would be cooler if you had 1 "see" the food and communicate to another "blind" agent to get it, that would show some fuzzy logic (keeping comms nonmathematical) and be interesting.

[d4sp1d3r]

The circuit is basic, equivalent to 50 or less transistors. The job is to figure how to connect input and output in a optimal fashion. Any addition in functionality, such as multi-agent cooperation, would further and unnecessarily complicate the implementation. Fuzzy controllers are the simplest, closest thing to human thinking you can shrink into a small number of components,? and easier to test and implement as well.

[sabriath]

I will agree with you on the simplest and easiest, but as far as being closest to human thinking, neural networks takes the cake on that one.....transistors are either on or off, neural networks are more analog which is closer to a neurons functionality. True that a neuron is an all-in output, but the transmitters can be weighted in inhibition or excitation as well as the gateways on the receiving synapse. Nice work though otherwise, good luck.

[d4sp1d3r]

The transistors are not an option, but imposed by the circuit model. I use neural networks for pattern recognition, which is its main usage. It requires more transistors to emulate a single neuron, furthermore that kind of connections are not easily implemented in digital circuits. This fuzzy controller was the best choice to use the limited number of connections in the best way possible. The genetic algorithm took care of adjusting the values for me.

[d4sp1d3r]

The link was added to the video description.

[sgwbutcher]

So such a group of rules is a fuzzy controller. I assume that what the author is saying when he says "optimal crispy boundaries" is that the GA optimized the parameters of the fuzzy sets...but he uses a lot of non-standard language and doesn't come right out and say what his fuzzy sets and fuzzy rules are, perhaps they're even encoded in a neural network which is a popular approach.

Some statements do make me scratch my head: "limited number of connections", I'm not sure what that might mean.

[sgwbutcher]

Using this "degree of truth" (membership in fuzzy sets) we can combine properties and their degree of truth into Fuzzy Rules: IF food is tasty and service is great THEN tip is large. This is a control rule, how much tip to leave. Fuzzy Logic enables you to determine the degree of truth of the antecedent and the degree of truth of the consequent (the implication) and combine the degree of truth for all rules with non-zero degrees of truth. Defuzzification then gives a crisp value of a tip.

[sgwbutcher]

What you describe are fuzzy sets. Typical examples are "short", "medium", "tall" for height or "cold", "cool", "warm", "hot" for temperature. Each of these sets can be described by full enumeration (ick) or a membership function of some sort that describes the degree to which a particular observation belongs to the fuzzy set. In your example, 6'0" might be 1.00 tall and 5'8" might be 0.87 tall. However, 5'8" might be 1.00 medium and 6'0" might be 0.9 medium.

[illustriouschin]

fuzzy logic is an artificial intelligence term as well, its been around for a long time. believe it or not some terms exist with different meanings depending on context!