Practical insights into fuzzy logic
Download as PPTX, PDF1 like133 views
This document discusses practical applications of fuzzy logic and provides examples of how fuzzy logic can be used to solve problems involving human concepts that are vague in nature. It describes two business case examples where fuzzy logic is used to determine a boat's location mode and a customer's trust index. The key benefits of fuzzy logic mentioned are that it provides a simple way to bridge human and machine reasoning by using linguistic rules and possibility distributions rather than crisp logic.
![Fuzzification & Defuzzification
Fuzzification Fuzzy
Inference
DeFuzzification
Fuzzy
Rulebase
Tolerance 1
Tolerance 2
pdf
pdf
pdf [0 … 1]](https://image.slidesharecdn.com/practicalinsightsintofuzzylogic-dileepa-181024020830/85/Practical-insights-into-fuzzy-logic-21-320.jpg)
Practical insights into fuzzy logic
- 1. Practical Insights into Fuzzy Logic Dileepa Jayathilake Manager Research – 99X Technology Cofounder - Tracified
- 2. A bit about me • Started with AI as an undergrad 16 years back • Developed Sinhala Text-to-Speech and Speech Recognition using Artificial Neural Networks and Hidden Markov Models • Authored the open source fuzzy logic library ALLib • Applied fuzzy logic in few industrial cases in Norway and China • More recently worked on Brain Wave Analysis • Current research interests: Blockchain, AI
- 3. What is Fuzzy Logic? • AI technique to use when you want to combine human concepts (which are inherently of vague nature) with computing • Solves problems where crisp logic fails
- 4. What is fuzzy logic good for today? Fuzzy Logic should be applied in cases where we have minimal knowledge about the process to be modelled - from a Fuzzy Logic Book from 90’s ?
- 5. Business Case 1: Geo-fencing Sailor wants the app to wake him up if the anchor is broken and the boat is being dragged by wind
- 6. Perfect World Solution Anchor Location Current Location Distance Moved (d) Did the boat move by more than a predefined distance?
- 7. Location Readings are not perfect Anchor Location Current Location Location readings are received along with a tolerance T1 T2 Distance Moved (d) Actual movement can be anything between d – T1 – T2 and d + T1 + T2
- 8. Tolerance can be Mild or Wild • High Accuracy Mode • Tolerance values are low • Readings are received frequently • Low Accuracy Mode • Tolerance values are high • Readings are received sparingly
- 9. It’s important to know which mode we are in • In high accuracy mode • Happily ignore intermittent readings with high tolerance • Wait for the reading with best accuracy and treat it as a perfect reading • In low accuracy mode • Try to make use of every reading as only few readings are received • Apply a geometric correction
- 10. Mode changes over time • Apply ‘mode recognition logic’ periodically • Lets say every 30 seconds • Then use the correct approach to find answer to the question “Should we ring the alarm?”
- 11. Determining the mode using Crisp Logic Example: If BEST_TOLERANCE < 20m && READING_COUNT >= 15 Mode = HIGH_ACCURACY Else Mode = LOW_ACCURACY Now consider following 3 cases Best Tolerance Reading Count Case 1 18 15 Case 2 21 15 Case 3 18 13 Only Case 1 qualifies as High Accuracy mode while intuition suggests that all 3 cases must be same mode
- 12. Alternative Can’t we write the logic in a form like: If BEST_TOLERANCE is LOW and READINGS_COUNT is VERY_HIGH Then there is a VERY_HIGH possibility that we are in HIGH_ACCURACY Mode Then we express our intuition on the case with further such rules: If BEST_TOLERANCE is MEDIUM and READINGS_COUNT is MORE_OR_LESS_HIGH Then there is a MEDIUM possibility that we are in HIGH_ACCURACY Mode If BEST_TOLERANCE is VERY_HIGH and READINGS_COUNT is LOW Then there is a LOW possibility that we are in HIGH_ACCURACY Mode
- 13. Bridging Human World and Machine World Humans think in terms of concepts Machines think in terms of numbers Can we build a logic to bridge these two worlds?
- 14. New form of the solution We express our intuition on the case as a Rulebase built out of concepts If BEST_TOLERANCE is LOW and READINGS_COUNT is VERY_HIGH Then there is a VERY_HIGH possibility that we are in HIGH_ACCURACY Mode If BEST_TOLERANCE is MEDIUM and READINGS_COUNT is MORE_OR_LESS_HIGH Then there is a MEDIUM possibility that we are in HIGH_ACCURACY Mode If BEST_TOLERANCE is VERY_HIGH and READINGS_COUNT is LOW Then there is a LOW possibility that we are in HIGH_ACCURACY Mode Then the machine solves the problem for us numerically! BEST_TOLERANCE = 12 && READINGS_COUNT = 18 Mode = HIGH_ACCURACY
- 15. Possibility: The key to bridge the two worlds Probability Likelihood Proportion Possibility Feasibility Ease of Attainment
- 16. Possibility Distribution Function Possibility Value X 1.0 0 Counterpart in the probability world is Probability Distribution Function
- 17. Membership Function Possibility Value Low (Reading Count) 1.0 0 Counterpart in the probability world is Singular Value 2 15
- 18. Meaning of X is LOW Possibility Value 1.0 0 2 15 • When we don’t have any more information about X, pdf for ‘X is LOW’ is the membership function of LOW • When we know X’s pdf, the pdf for ‘X is LOW’ is mf(LOW) o pdf(X)
- 19. Fuzzy Inference Output pdf = ALSO{(input AND inputrulebase)IF-THEN outputrulebase}
- 20. Solution Configuration Tolerance 1 = 8m ; Tolerance 2 = 15m ; Reading Count = 12 Is High Accuracy Mode?
- 21. Fuzzification & Defuzzification Fuzzification Fuzzy Inference DeFuzzification Fuzzy Rulebase Tolerance 1 Tolerance 2 pdf pdf pdf [0 … 1]
- 22. Challenges • Defining membership functions • Incompleteness of rulebase
- 23. Business Case 2: Customer Trust Index How to quantify business value of a customer when you have sales data over a period? Quantity : Total amount a customer has spent Regularity : Frequency of Purchasing Consistency : Predictability of the customer
- 24. Direct Solution Parameter High Score Low Score Quantity 10 1 Regularity 5 3 Predictability 8 2 E.g. Score(High Quantity, Low Regularity, Low Predictability) = 10 * 3 * 2 = 60
- 25. Fuzzy Logic based Solution Define fuzzy input variables Quantity, Regularity and Predictability Define fuzzy sets for each input variable’s domain High_Value, Medium_Value, Somewhat_Low_Value, Insignificant_Value, …,Frequent, Regular, Occational, …, Recurring, Quasi-Predictable, Adhoc, … Define fuzzy output variable Trust Index Define fuzzy sets for output variable’s domain Regular_Customer, Experimentor, Endorser, Adhoc_Customer, … Define rulebase If (Quantity is Somewhat_Low_Value) AND (Regularity is Frequent) Then Trust Index is Experimentor If (Quantity is High_Value) AND (Regularity is Regular) AND (Predictability is Quasi-Predictable) Then Trust Index is Endorser Ask the question: If Quantity = 100 ; Regularity = 7 ; Predictability = 8, What is the Trust Index?
- 26. Business Case 3: Customer Pain Point Analysis How to quantify the importance of customer pain points reported?
- 27. Facts & Ideas • Pain points are reported in 3 ways • Phone complaints that result in a refund • Phone complaints that do not result in a refund • Internet reviews with low ratings • Intuition suggests that the 3 types are in decreasing order of importance • Pain points that have an association of loss in customer trust index have a special importance
- 29. Direct Solution * alpha and beta are selected to be 2.0 and 1.5 to reflect the relative importance of the factor they are multiplied with
- 30. Fuzzy Logic based Solution • Input variables • Customer Trust Index • Drop in Customer Trust Index • Phone complaint causing refund • Phone complaint not causing refund • Internet Review • Output variable: Pain point score • Multiplex before sending to fuzzy inference • Evaluate per-review, defuzzify and take the sum
- 31. Results
- 32. Advantages using Fuzzy Logic • Simplicity • Ready for realtime applications (computationally lean) • Human readability of rulebase • Rulebase doesn’t have to be exhaustive • Ability to explain the result
- 33. Blockchain AI • Blockchains already use AI based agents as oracles • Real “Blockchain AI” would be to use AI in consensus • Fuzzy logic, because of its simplicity, is a good contender for that
- 34. Checkout my Fuzzy Logic library in Github
- 35. Questions