Artificial Intelligence: Search, Bayes and Evaluation Metrics
What this section tests
The Artificial Intelligence section tests conceptual and quantitative fluency rather than coding: search strategies, propositional logic, probability and Bayes' rule, basic linear algebra and optimisation, supervised learning fundamentals, and evaluation metrics.
Core rules
- BFS is complete and optimal for unit costs but uses O(b^d) memory; DFS is memory-light but not optimal; uniform-cost search expands by lowest path cost.
- A* is optimal with an admissible heuristic (one that never overestimates); consistency implies admissibility.
- Bayes: P(A|B) = P(B|A)P(A)/P(B). The base rate usually dominates the posterior.
- Bias–variance: high bias underfits, high variance overfits; regularisation and more data reduce variance.
- Precision = TP/(TP+FP); recall = TP/(TP+FN). They trade off as the decision threshold moves.
- Gradient descent updates θ ← θ − α∇J(θ); the learning rate α sets step size, not direction.
Worked example
A classifier is tested on 200 cases: TP = 40, FP = 10, FN = 20, TN = 130. Precision = 40/50 = 0.80. Recall = 40/60 = 0.667. Accuracy = (40+130)/200 = 0.85. Now the trap: predicting "negative" for every case scores 130/200 = 0.65 accuracy while detecting nothing at all. Accuracy alone is a poor summary under class imbalance, which is why precision, recall and F1 are reported instead.
Traps
Swapping precision and recall denominators; quoting accuracy on a 95:5 class split; assuming a heuristic is admissible because it seems close; treating correlation in a dataset as evidence of a causal mechanism.