Expert Systems & Fuzzy Logic: Rules, Uncertainty & Real-World AI

Core Concepts

Expert Systems: The Digital Consultant

Definition: An expert system is a rule-based AI program that mimics human expertise in a narrow domain (e.g., medicine, finance) by combining:

• A knowledge base (facts + rules)
• An inference engine (reasoning logic)
• A user interface (question/answer)

classDiagram
    class ExpertSystem {
        +KnowledgeBase
        +InferenceEngine
        +UserInterface
    }
    class KnowledgeBase {
        +Facts
        +Rules
    }
    class InferenceEngine {
        +ForwardChaining
        +BackwardChaining
    }
    ExpertSystem --> KnowledgeBase : contains
    ExpertSystem --> InferenceEngine : uses
    ExpertSystem --> UserInterface : interacts

How It Works:

1. Knowledge Acquisition: Experts encode their expertise as rules (IF-THEN statements).
2. Inference: The engine applies rules to facts to deduce new conclusions.
3. Explanation: Justifies its reasoning (critical for trust).

Example: Loan Approval System (like Nepal’s NMB Bank)

• Rule: IF (income > 50k AND credit_score > 700) THEN approve_loan(YES)
• Facts: income = 60k, credit_score = 750
• Inference: Engine fires the rule → approve_loan(YES)

Fuzzy Logic: Handling the Gray Areas

Definition: A mathematical framework to represent uncertainty where truth values range between 0 (false) and 1 (true), unlike binary logic (0/1). Used when data is imprecise (e.g., "tall," "hot").

Membership Functions: The Heart of Fuzzy Logic

For a fuzzy set Height = {short, medium, tall} over domain X = {150, 160, 170, 180, 190}:

• Short: Membership increases linearly from 150 (0) to 170 (1).
• Medium: Overlaps with "short" and "tall" (e.g., 170 = 0.5 for medium).
• Tall: Membership increases from 170 (0) to 190 (1).

Worked Example: Air Conditioner Temperature Control

• Fuzzy Sets: Cold = {20-24°C}, Comfortable = {22-26°C}, Hot = {24-30°C}
• Rule: IF temperature IS Hot THEN set_fan_speed(High)
• Input: 28°C → Membership in Hot = 0.8, Comfortable = 0.2
• Defuzzification: Combine outputs (e.g., weighted average) → fan_speed = 75%

In the Real World

1. eSewa (Nepal):

   • Expert System: Uses rule-based logic to validate bill payments (e.g., IF (meter_number IN database AND amount > 0) THEN process_payment).
   • Fuzzy Logic: Adjusts "congestion fees" for delayed payments based on vague rules like "slightly late" (1–3 days) vs. "very late" (>7 days).

2. Pathao (Ride-Hailing):

   • Dynamic Rule Base: Adjusts surge pricing using fuzzy logic for demand (e.g., IF (wait_time > 10 mins AND drivers < 5) THEN surge_multiplier = 1.5).

3. NTC Traffic Management:

   • Fuzzy Controller: Regulates traffic lights at busy intersections using rules like: IF (car_queue > 20 AND pedestrian_queue = 0) THEN green_light_duration = 45s.

Development Phases of an Expert System

flowchart LR
    A["Identify Domain"] --> B["Knowledge Acquisition"]
    B --> C["Knowledge Representation"]
    C --> D["Design Inference Engine"]
    D --> E["Build User Interface"]
    E --> F["Testing & Validation"]
    F --> G["Deployment & Maintenance"]

Key Steps:

1. Domain Identification: Narrow the scope (e.g., "diabetes diagnosis" vs. "general medicine").
2. Knowledge Acquisition: Interview experts or extract rules from data.
3. Knowledge Representation: Use production rules (IF-THEN), semantic networks, or frames.
4. Inference Engine: Choose forward (data-driven) or backward (goal-driven) chaining.
5. Validation: Test with real-world cases (e.g., 100 patient records for a medical system).

Comparison: Rule-Based vs. Machine Learning

When to Use Rule-Based:

• Static Environments: Tax calculation, grammar checkers.
• High Stakes: Medical diagnosis (explainability matters).
• Limited Data: Few examples but clear rules.

When to Use ML:

• Dynamic Environments: Stock trading, social media trends.
• Unstructured Data: Images, text (NLP).

Fuzzy Logic Operators

Fuzzy logic extends Boolean operators (AND, OR, NOT) to handle partial truths:

• AND: min(a, b)
• OR: max(a, b)
• NOT: 1 - a

Example: Washing Machine Logic

• Rule: IF (dirt_level = High AND water_hardness = Medium) THEN detergent_amount = 70%
• Calculation:
  • dirt_level = 0.9, water_hardness = 0.6
  • AND result = min(0.9, 0.6) = 0.6
  • detergent_amount = 0.6 * 100% = 60% (rounded to 70% per rule).

Worked Example: Covid-19 Vaccine Recommender (PEAS Framework)

Performance: Recommend vaccines based on age, health conditions, and side effects. Environment: Dynamic (new variants, drug interactions). Actuators: Database queries, user alerts. Sensors: Patient records, CDC guidelines.

flowchart TD
    A["Patient Data"] --> B["Rule Base"]
    B --> C["Inference Engine"]
    C --> D["Recommendation"]
    D --> E["User Alert"]
    B -->|"Example Rule"| F["IF (age > 65 AND has_heart_condition) THEN recommend_Pfizer"]

Fuzzy Rule Example:

• Rule: IF (risk_level = High AND mobility = Low) THEN priority = Urgent
• Membership:
  • risk_level = 0.8 (from symptoms + age)
  • mobility = 0.3 (uses wheelchair)
  • AND result = min(0.8, 0.3) = 0.3
  • Defuzzification: Combine with other rules → priority = "High".

Exam Tip

1. Define Clearly:

   • Expert system = "AI system that encodes human expertise as rules."
   • Fuzzy logic = "Logic that handles partial truth values (0–1)."

2. Draw Diagrams:

   • Always sketch a rule-based system flowchart (knowledge base → inference engine → output).
   • For fuzzy logic, plot membership functions (triangular/trapezoidal).

3. Link to Nepal:

   • eSewa: Rule-based for bill validation.
   • NTC Traffic: Fuzzy logic for signal timing.
   • Banks: Loan approval rules (e.g., IF (income < 30k AND loan > 2M) THEN reject).

4. Common Pitfalls:

   • Forgetting to defuzzify (convert fuzzy outputs to crisp actions).
   • Confusing forward chaining (data → conclusion) with backward chaining (goal → data).

5. Past Exam Patterns:

   • 20%: Define + example (e.g., "Expert system for Covid prediction").
   • 30%: Design a rule base or fuzzy set (use realistic domains like traffic or loans).
   • 25%: PEAS framework (describe performance, environment, actuators, sensors).
   • 25%: Convert sentences to FOPL (First-Order Predicate Logic) or explain inference.