Question Pattern & Exam Analysis

1. Course Syllabus Overview

The official syllabus covers the following primary domain areas:

---

2. Topic Frequency & Marks Heatmap

---

3. Question Composition & Weightage Balance

• ~60% Numerical & Step-by-Step Algorithmic Tracing: High-scoring, deterministic problems (e.g., Value Iteration, Fuzzy Inference, Softmax/Cross-Entropy, FOPL Resolution, 4-Queens Forward Checking, A*/Hill-Climbing on 8-puzzle, Minimax/Alpha-Beta on game trees).
• ~30% Comparative & Analytical Logic: Algorithm comparisons, theoretical proofs, trade-off evaluations, and performance analyses.
• ~10% Conceptual & Real-World Case Studies: System design choices, ethics in autonomous systems, and agent framework formulations.

---

4. Detailed Analysis of Core Question Patterns

Pattern A: Carbon-Copy & Direct Repeated Questions Across Years

Several questions have appeared virtually identically across multiple exam years:

1. FOPL Resolution Proof ("Jack and Curiosity Killed the Cat"):
   • Appeared in: 2023 Question 4(a) and 2024 Question 2(a) — Identical premise sentences (A through G), background knowledge, and goal.
   • Required Steps:
     1. Convert FOPL statements into Conjunctive Normal Form (CNF).
     2. Standardize variables and apply Skolemization where required.
     3. Draw the step-by-step Resolution Graph deriving the empty clause ($\square$).
2. Fuzzy Logic Inference Tracing:
   • Appeared in: 2023 Question 4(b) (Driving Speed based on Cloud Cover & Temp) and 2024 Question 2(b) (Repair Utilization based on Delay & Servers).
   • Required Steps: Fuzzification of crisp inputs $\to$ Mamdani Min-Rule Evaluation $\to$ Max-Aggregation $\to$ Center of Area (COA) Defuzzification calculation.
3. Softmax Probabilities & Cross-Entropy Loss Calculation:
   • Appeared in: 2023 Question 3(a) (Logits [5.5, 2.9, 3.2, 0.8]) and 2024 Question 7(b) (Logits [0.2, 2.5, 0.3]).
   • Required Steps: Calculate $P(y_i) = \frac{e^{z_i}}{\sum_j e^{z_j}}$, represent true label as a one-hot vector, and compute loss $L = -\ln(P_{\text{true}})$.
4. Bayesian Posterior Probability in Diagnostic Problems:
   • Appeared in: 2022 Question 5(b) (Starfleet androids), 2023 Question 7(b) (Meningitis & Stiff neck), and 2024 Question 6(a) (Disease incidence 1%, TP 99%, TN 95%).
   • Required Formula: Apply Bayes Rule $P(D \mid T) = \frac{P(T \mid D)P(D)}{P(T \mid D)P(D) + P(T \mid \sim D)P(\sim D)}$.
5. 8-Puzzle Search Tree Execution:
   • Appeared in: 2022 Question 2(b) (Hill Climbing 4 iterations) and 2024 Question 1(c) (A* path tree expansion with misplaced tiles heuristic).
6. MCTS Rollouts & Strategy Extraction:
   • Appeared in: 2023 Question 2(b) and Incourse 2023 Question 2(c) — Given initial tree stats ($t_i, n_i$), performing 2 rollout iterations, updating visit/reward counts via UCB1, and extracting optimal strategy.
7. MRV & LCV Heuristic Rationale:
   • Appeared in: Incourse 2023 Question 1(c) and 2024 Question 3(b) — Why selecting the Most Constrained Variable (MRV) fails fast while the Least Constraining Value (LCV) maximizes remaining search space.

---

Pattern B: Numerical & Algorithmic Tracing Problems (~60% of Marks)

These problems require step-by-step mathematical work:

• Game Search & MCTS:
  • Minimax & $\alpha$-$\beta$ Pruning: Labeling node values and explicitly identifying pruned subtrees using $\alpha \ge \beta$ conditions (2022 Q5c, 2024 Q5).
  • MCTS Rollouts & Strategy: Given initial tree stats ($t_i, n_i$), performing 2 rollout iterations, updating visit/reward counts, and extracting optimal strategy (2023 Q2b, Incourse 2023 Q2c).
• CSP Search:
  • Forward Checking & Backtracking: Tracing a 4-Queens search tree with dynamic domain reduction tables (2024 Q3a).
  • Arc Consistency ($AC-3$): Detecting inconsistency in binary relational arcs ($v(N,W)$) and eliminating domain elements (2022 Q3a).
• Reinforcement Learning & MDPs:
  • Value Iteration Tracing: Computing state utilities iteratively using the Bellman backup equation: $V_{k+1}(s) = \max_a \sum_{s'} T(s,a,s') \left[ R(s) + \gamma V_k(s') \right]$ (2024 Q4a).
  • Q-Learning Table Updates: Updating Q-table values for specified state-action transitions using: $Q(s,a) \leftarrow Q(s,a) + \alpha \left[ r + \gamma \max_{a'} Q(s',a') - Q(s,a) \right]$ (2023 Q5b).
• Machine Learning & Neural Networks Tracing:
  • ID3 Decision Trees: Computing baseline Entropy $H(S) = -\sum p_i \log_2 p_i$, attribute Information Gain $Gain(S,A)$, and selecting the root split (2022 Q6b).
  • Perceptron Weight Updates: Calculating linear combiner output $v = \mathbf{w}^T \mathbf{x}$ and applying weight updates $\Delta w_i = \eta (t - y) x_i$ (2022 Q7a).
  • Softmax & Loss: Computing Softmax output vectors and cross-entropy loss values (2023 Q3a, 2024 Q7b).
• Probability:
  • Bayes Theorem: Medical diagnosis style problems with prior, likelihood, and posterior computation (2022 Q5b, 2023 Q7b, 2024 Q6a).
  • Joint Probability Queries: Marginal and conditional probability extraction from joint distribution tables (2023 Q7c).

---

Pattern C: Mathematical Proofs & Formal Derivations (~10% of Marks)

• Consistency Implies Admissibility: Proving that if a heuristic $h(n)$ satisfies consistency ($h(n) \le c(n,a,n') + h(n')$), it must also be admissible ($h(n) \le h^*(n)$) (2024 Q1a).
• Search Algorithm Equivalence Proofs / Counterexamples:
  • Proof that $BFS$ is a special case of $UCS$ when all step costs are equal ($c=1$).
  • Proof that $DFS$ is a special case of Best-First search when $h(n) = -\text{depth}(n)$.
  • Proof that $UCS$ is a special case of $A^*$ search when $h(n) = 0$ (2023 Q1c).
• Markov Property Formalization: Mathematically defining $P(S_{t+1} \mid S_t, S_{t-1}, \dots, S_0) = P(S_{t+1} \mid S_t)$ and analyzing its impact on transition matrix size (2023 Q5a, 2024 Q6b).

---

Pattern D: Algorithmic Comparative & Analytical Questions (~20% of Marks)

• MCTS vs $\alpha$-$\beta$ Pruning: Evaluating performance on huge branching factors (e.g., Go) where full heuristic evaluation functions fail (2023 Q2b, Incourse 2023 Q2d).
• Value Iteration vs Q-Learning: Comparing Model-based (requires transition model $T$ and reward model $R$) vs Model-free (learns directly from experience tuples $\langle s,a,r,s'\rangle$) (2024 Q4b).
• Perceptron Learning Rule vs Delta Rule: Comparing binary step activation vs differentiable continuous activation (Sigmoid/Linear) in handling non-linearly separable data (2023 Q3a).
• Activation Function Failures:
  1. Sigmoid: Vanishing gradient problem in deep layers due to small derivative values ($\phi'(v) \le 0.25$).
  2. ReLU: "Dying ReLU" problem where negative inputs produce zero gradient permanently.
  3. Solutions: Leaky ReLU ($\phi(x) = \max(\alpha x, x)$) or ELU (2024 Q7a).
• CSP Heuristics (MRV vs LCV): Why selecting the Most Constrained Variable (MRV) fails fast, while selecting the Least Constraining Value (LCV) maximizes the remaining search space (2023 Incourse Q1c, 2024 Q3b).

---

Pattern E: Real-World Scenarios & System Design (~10% of Marks)

• Autonomous Driving Ethics: Addressing ethical dilemmas, safety guarantees, and risk minimization in autonomous vehicles (DhakaTruck case study) (2022 Q1a).
• Life-Critical vs Time-Sensitive Navigation: Comparing A*, Greedy Best-First, Dijkstra, and Weighted A* for emergency medical drone delivery vs city food delivery (Incourse 2024 Q1).
• Medical Diagnostic Advisor Design: Choosing between Crisp Rule-Based System, Backward-Chaining Expert System, and Fuzzy Inference System based on uncertainty and clinical workflow (Incourse 2023 Q2a).

---

5. Complete Question-to-Topic Mapping

Final Exam 2022 (25th Batch)

---

Final Exam 2023 (26th Batch)

---

Final Exam 2024 (27th Batch)

---

In-Course Exam 2023 (26th Batch)

In-Course Exam 2024 (27th Batch)

---

6. Strategic Preparation & High-Yield Topic Roadmap

To maximize your score on the exam, organize your revision using the following priority tiers:

Tier 1: Guaranteed Full-Mark Scoring Topics (Highest ROI)

These topics feature deterministic algorithms with fixed step-by-step recipes and appear every single year:

1. FOPL Resolution: Practice CNF conversion (Skolemization, variable standardization) and resolution graph derivations — the "Jack killed the cat" problem appeared verbatim in both 2023 and 2024.
2. Fuzzy Inference Systems: Practice 4-step Mamdani calculations (Fuzzification, Min, Max, COA Defuzzification).
3. Softmax & Cross-Entropy: Practice calculating Softmax probabilities from logits and natural log loss from one-hot encoded true labels.
4. Bayesian Posterior Probability: Practice Bayes theorem medical diagnosis problems — appeared in all 3 years with different scenarios.
5. MDP Value Iteration: Practice tabular utility updates using Bellman backups with discount factor.
6. Q-Learning Updates: Practice Q-table updates with learning rate $\alpha$ and $\epsilon$-greedy policy analysis.

Tier 2: Essential Tree Search & Algorithmic Tracing

1. Game Search: Minimax tree labeling, $\alpha$-$\beta$ pruning bounds (identify all pruned nodes), MCTS rollout tracing with UCB1.
2. CSP 4-Queens: Forward checking table updates, domain reduction, and backtracking trigger conditions.
3. 8-Puzzle Search: A* state expansion with $f(n) = g(n) + h(n)$ and Hill-Climbing state evaluations.
4. ID3 Decision Trees: Baseline entropy and Information Gain calculations.

Tier 3: Core Theoretical Derivations & Analytical Rationales

1. Heuristic consistency $\implies$ admissibility proof (triangle inequality induction).
2. Search algorithm reduction/equivalence proofs ($BFS \subset UCS \subset A^*$).
3. Activation function gradient analysis (Sigmoid vanishing → ReLU dying → Leaky ReLU solution).
4. Hidden Markov Models — 3 fundamental problems: Evaluation (Forward Algorithm), Decoding (Viterbi), Learning (Baum-Welch).
5. Markov Property formalization and its impact on MDP complexity.
6. MRV & LCV heuristic rationale for CSP search efficiency.