Beyond the Hype: Why ML Exists — and Why Search Algorithms Still Matter
Introduction Everyone talks about Machine Learning like it appeared out of thin air. But the truth is, ML didn't come first. Before neural networks and gradient descent, programmers tried to build intelligent systems the "obvious" way — by writing rules. Lots and lots of rules. This post tells two stories:
Why rule-based systems eventually broke down — and why ML was the necessary answer. Why Search Algorithms & Planning, one of AI's oldest branches, still powers the world quietly in the background.
Part A: Why Machine Learning? The Limits of Rule-Based Systems The "Just Write the Rules" Era In the early days of AI, the dominant approach was simple in theory: hire domain experts, ask them how they make decisions, and encode that knowledge into if-else logic. These were called Expert Systems. A spam filter from the early 2000s looked something like this: pythondef is_spam(email): if "FREE MONEY" in email: return True if "CLICK HERE NOW" in email: return True if "CONGRATULATIONS YOU WON" in email: return True return False Simple. Clean. And for a while — it worked. Where Rules Break Down: The Spam Example Let's think about what happens next. Spammers are smart. Within weeks of a new rule being deployed, they adapt:
Now your rule-based filter is outdated the moment it ships. Every time spammers evolve, a human engineer has to manually update the rules. This is not sustainable. The core problems with rule-based systems: ProblemWhat It MeansBrittlenessOne edge case can break the entire systemScalabilityYou can't write rules for every possible scenarioMaintenanceRules need constant manual updatingNo generalizationRules only handle exactly what they were written forExpert bottleneckYou need a domain expert available at all times A Real-World Failure: Medical Diagnosis In the 1970s and 80s, researchers built MYCIN — a rule-based expert system for diagnosing bacterial infections. It had over 600 hand-crafted rules written by doctors. MYCIN worked reasonably well in controlled settings. But it had a critical flaw: it couldn't handle uncertainty. Real patients don't fit neatly into predefined categories. A symptom that appears in one patient might mean something completely different in another. When doctors tried to extend MYCIN to new diseases, they had to start from scratch — writing hundreds of new rules by hand. It simply didn't scale. How Machine Learning Solves This Instead of writing rules, ML takes a different approach: show the system thousands of examples and let it find the patterns itself. Rule-Based: Human writes rules → System applies rules ML: Human provides data → System learns rules A spam filter trained on 100,000 labeled emails doesn't need you to tell it what spam looks like. It learns the statistical patterns on its own — and when spam evolves, you just retrain with new data. This shift — from programming behavior to learning behavior — is what made ML a necessary evolution, not just a trendy upgrade.
Part B: Search Algorithms & Planning — The Unsung Hero of AI What Is Search & Planning? While everyone's talking about neural networks, a huge chunk of the AI-powered world runs on something much older: Search Algorithms and Planning. At its core, Search & Planning is about finding the best sequence of actions to reach a goal from a starting point. It works by exploring a state space — all possible configurations of a problem — and finding an optimal path through it. Think of it like a chess player thinking several moves ahead: Current State → Possible Actions → Next States → ... → Goal State The Classic Example: Pathfinding The most famous search algorithm is *A (A-star)**. It's used to find the shortest path between two points, and it's everywhere:
Google Maps routing Game AI (enemy characters finding their way to you) Robot navigation Logistics & delivery route optimization
Here's the intuition behind A*: f(n) = g(n) + h(n)
Where: g(n) = actual cost from start to current node h(n) = estimated cost from current node to goal (heuristic) f(n) = total estimated cost By combining actual cost with a smart estimate, A* finds the optimal path far more efficiently than blindly searching every possibility. A Visual: State Space Search START | ├── Action A → State 1 | ├── Action C → State 3 ✓ GOAL | └── Action D → State 4 | └── Action B → State 2 └── Action E → State 5 The algorithm explores this tree smartly, pruning paths that are clearly worse, until it reaches the goal. Where Search & Planning Is Used TODAY You might think this is ancient history. It's not.
GPS Navigation Every time Google Maps calculates your route, it's running a variant of Dijkstra's or A* on a graph of millions of nodes. Pure ML cannot do this reliably — you need exact, optimal paths.
Game Development Every NPC (non-player character) that chases you or avoids obstacles uses pathfinding. Games like The Last of Us and Halo use sophisticated planning systems for enemy AI behavior.
Robotics Boston Dynamics robots use motion planning algorithms to decide how to move limbs in 3D space. Pure ML would be too unpredictable for physical safety.
Logistics & Supply Chain Amazon's warehouse robots use planning algorithms to optimize which route to take when picking items — saving millions of dollars in efficiency.
AlphaGo (Yes, even Deep Learning uses Search) Google DeepMind's AlphaGo — which beat the world's best Go player — combined deep learning with Monte Carlo Tree Search (MCTS), a planning algorithm. Neither alone was sufficient. Why It Still Matters in the Age of Deep Learning Deep Learning is powerful, but it has weaknesses:
It needs massive amounts of data It's hard to guarantee correctness (a neural network might give a wrong answer confidently) It's computationally expensive for simple planning tasks
Search & Planning algorithms are:
Exact — they provably find optimal solutions Interpretable — you can trace every decision Efficient for well-defined problems Combinable with ML (as AlphaGo proved)
The future of AI isn't ML replacing everything — it's ML and classical AI working together.
Conclusion
The history of AI is not a straight line from "dumb rules" to "smart neural networks." It's a rich ecosystem where different approaches solve different problems. Rule-based systems taught us what computers could do — and showed us their limits. Machine Learning broke through those limits by learning from data instead of following instructions. And Search & Planning has been quietly powering some of the most critical AI applications in the world — from your GPS to warehouse robots to AlphaGo. Understanding all of these branches doesn't just make you a better AI student. It makes you a better engineer — one who knows which tool to reach for, and why.
Written by [Your Name] | AI Assignment 2 | Spring 2025 Special thanks to our instructor Bilal Jan and TA Raqeeb for their guidance.
Published as part of the AI Assignment 2 Series | Spring 2026
References:
Russell, S. & Norvig, P. (2020). Artificial Intelligence: A Modern Approach (4th ed.). Pearson. Shortliffe, E.H. (1976). MYCIN: Computer-Based Medical Consultations. Elsevier. Hart, P., Nilsson, N., & Raphael, B. (1968). A formal basis for the heuristic determination of minimum cost paths. IEEE Transactions on Systems Science and Cybernetics. Silver, D. et al. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529, 484–489. Géron, A. (2022). Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow. O'Reilly.