Skip to content
Tutorials Center
My learning

Topics

Lesson 1 of 46 ยท Intelligent Agents

What Is Classical AI?

Understand classical AI: how symbolic, statistical and deep learning approaches differ, a short history with AI winters, and what search, logic and planning still do best.

When people say "AI" today, they usually mean a chatbot or an image generator trained on enormous piles of data. Those systems are impressive, but they are only one branch of a field that is more than seventy years old. Classical AI is the other big branch: programs that reach goals by searching through possibilities, reasoning with explicit rules, planning sequences of actions and weighing uncertainty with probability. Nothing is learned from millions of examples. Instead, a human describes the problem precisely, and an algorithm solves it.

This course teaches that branch from the ground up, and every algorithm runs live in your browser in Python. You will solve puzzles, schedule resources, beat people at board games, prove theorems, plan robot motions and track a robot's position through noisy sensor readings. These ideas are inside route planners, game engines, compilers, chip-design tools, airline schedulers, spacecraft, warehouse robots and, increasingly, inside modern machine-learning systems that use search on top of learned models.

Three families of AI

It helps to put the whole field on one map. There are three broad ways to make a machine behave intelligently.

Approach Where the "knowledge" lives Typical tools Strength Weakness
Symbolic (classical) Hand-written rules, models and goals Search, logic, planning, constraint solving Exact, explainable, no training data Needs a precise model of the problem
Statistical Parameters fitted to data Regression, decision trees, Bayesian models Works with noisy data and few assumptions Needs good features and labelled data
Deep learning Millions or billions of learned weights Neural networks, transformers Handles images, speech and language Opaque, data-hungry, can be confidently wrong

The families are not rivals so much as tools for different jobs. If you can write down the rules of a problem (chess, a delivery schedule, a Sudoku), a symbolic solver can often find a provably correct or optimal answer. If you cannot write the rules down (is this photo a cat?), you learn them from examples.

graph TD
    AI["Artificial intelligence"] --> SYM["Symbolic / classical<br/>search, logic, planning"]
    AI --> ML["Machine learning"]
    ML --> STAT["Statistical learning<br/>regression, trees"]
    ML --> DL["Deep learning<br/>neural networks"]
    SYM -.-> HYB["Hybrid systems<br/>learned model + search"]
    DL -.-> HYB

AI is a family tree, and the most successful modern systems often combine branches.

Rules versus learning in a few lines of code

The difference is easiest to see on a toy task: deciding whether a message is spam. A symbolic programmer writes the rule. A statistical programmer lets the data pick the rule. Run both and compare.

Python A hand-written rule versus a learned rule

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Both classifiers get every training message right, but they disagree on the new message (3 links, no exclamation marks): the rule says spam, the learned threshold of 4 says no. Neither is "wrong" without more data. The rule is readable and you can argue with it; the learned threshold adapts if the data changes. Classical AI sits firmly on the "readable and argue with it" side: you can inspect why the answer came out the way it did, and often prove it is correct.

A short history

The story has a rhythm of big promises followed by sobering reality checks.

  • 1950s. Alan Turing asks whether machines can think. In 1956 a summer workshop at Dartmouth gives the field its name. Early programs prove logic theorems and play checkers.
  • 1960s. Search becomes the central idea. The A* algorithm (1968) and early robots such as Shakey, which planned its own moves, appear. Optimism is huge: some researchers predict human-level AI within a generation.
  • 1970s: first AI winter. Problems turn out to be much harder than toy versions. Combinatorial explosion, tiny computers and critical reports cut funding.
  • 1980s. Expert systems encode the knowledge of human specialists as if-then rules, and companies buy them. Neural networks also revive.
  • Late 1980s to 1990s: second AI winter. Expert systems are brittle and expensive to maintain; specialised hardware collapses commercially. At the same time, probabilistic methods such as Bayesian networks and hidden Markov models quietly take over speech and robotics.
  • 1997. Deep Blue defeats the world chess champion using fast alpha-beta search and a hand-tuned evaluation function. You will build the same ingredients in this course.
  • 2010s onward. Large datasets, GPUs and better training methods make deep learning dominate perception and language. Systems such as AlphaGo combine a learned network with Monte Carlo tree search, a perfect example of a hybrid.

An AI winter is a period when hype outruns results and funding dries up. The lesson classical AI learned from the winters is to be honest about what a method needs: a good model of the world, enough computing power, and a problem small enough, or structured enough, to be tractable.

Note: "Classical" does not mean "obsolete". Search, constraint solving and planning were not replaced by neural networks. They moved into places where correctness and guarantees matter, and they are paired with learning where that helps.

Can a machine think? The Turing test and its critics

In 1950 Alan Turing replaced the vague question "can machines think?" with a practical one, the imitation game, now called the Turing test. A human judge chats by text with two hidden partners, one human and one machine. If the judge cannot reliably tell which is which, the machine is said to pass. The test needs no definition of "thinking"; it only asks whether behaviour is indistinguishable from a person's.

Criteria for a fair test include a text-only channel (so appearance and voice give nothing away), a judge who is free to ask anything, a fixed time limit and many judges, since a single lucky guess means little. Chatbots such as ELIZA (1966) fooled some people with simple pattern matching, and modern language models can hold long, convincing conversations, which shows the weakness of the test: it rewards fooling the judge, not understanding.

The best-known objection is the Chinese Room argument (John Searle, 1980). Imagine a person who knows no Chinese sitting in a room with a huge rulebook. Slips of Chinese characters come in, and the person looks up the matching reply in the book and passes it out. To the people outside, the room "speaks Chinese", yet nobody inside understands a word. Searle argued that a program that only manipulates symbols by rule has syntax but no meaning, however good its answers look. Replies to the argument say the whole system (person plus book) understands, or that a robot grounded in the real world would. The debate is still open.

Because open chat is easy to bluff, researchers proposed harder tests. The Winograd Schema Challenge asks questions where one word flips the answer and commonsense is needed: in "The trophy does not fit in the suitcase because it is too big", it is the trophy, but with "too small" it is the suitcase. Today such benchmarks are used to measure language and reasoning ability rather than to settle whether a machine "thinks".

Types of AI

Two tidy classifications appear in almost every introduction.

By capability (how general the intelligence is):

Type Meaning Status
Narrow AI (weak AI) Does one task, or a small family of tasks, very well All real systems today: chess engines, route planners, spam filters, recommendation engines
General AI (AGI, strong AI) Matches a human across essentially any intellectual task, learns new skills and transfers knowledge Research goal; not achieved
Super AI (ASI) Surpasses the best humans in every field Hypothetical

By functionality (what the system can do with its environment):

Type Meaning Example
Reactive machines Respond to the current input; no memory Deep Blue
Limited memory Use recent observations to decide Self-driving cars tracking nearby vehicles
Theory of mind Would model the beliefs and intentions of others Research stage
Self-aware Would have consciousness of its own state Hypothetical

The first two functional types map onto the agent designs in Agents and Environments: a reactive machine is a simple reflex agent, and an agent that keeps internal state has limited memory. A third axis divides approaches rather than systems: symbolic AI (also called good old-fashioned AI) represents knowledge as symbols and rules and reasons by manipulating them, which is the subject of this course, while connectionist AI stores knowledge in the weights of neural networks. As the table earlier in this lesson showed, the two are complements, and hybrids are common.

What classical methods still do best

Reach for a classical technique when:

  • The rules are known and exact. Board games, scheduling, configuration, logistics, circuit verification.
  • You need a guarantee. "This plan is the cheapest possible" or "this schedule violates no constraint" are things search and constraint solvers can promise.
  • You need an explanation. A planner can print the sequence of steps it chose; a theorem prover can print its proof.
  • Data is scarce. A solver needs a model, not a million labelled examples.
  • Safety matters. Controllers for robots, aircraft and medical devices are often built from verified, explainable parts.

Their main weakness is the model itself. If you cannot describe the problem in states, actions and constraints, or the world is too messy to model exactly, learning-based methods are the better starting point. A big theme of the last chapters is how probability bridges that gap by letting a classical reasoner admit it is uncertain.

How this course is organized

The chapters follow a path from simple to rich settings:

  1. Intelligent agents (this chapter): the vocabulary used everywhere else.
  2. Uninformed and informed search: finding action sequences by exploring, then exploring smartly with heuristics.
  3. Constraint satisfaction: when the answer is an assignment of values, not a path.
  4. Adversarial search: planning against an opponent.
  5. Logic: representing knowledge and deriving new facts from it.
  6. Planning: search over rich, structured descriptions of actions.
  7. Probabilistic reasoning: acting under uncertainty.
  8. Robotics: connecting algorithms to bodies, sensors and motion.
  9. Projects and practice: four larger programs, exercises, interview questions and a cheat sheet.

You only need basic Python. Where a lesson leans on data structures, it links to the matching lesson in the DSA tutorial, such as queues, priority queues and graph BFS. Every code box is editable: change a number, press Run and see what happens. Learning AI algorithms is mostly learning to predict what they will do, so run experiments often.

A tool chooser

As a last warm-up, here is the kind of decision you will learn to make with confidence. The function below maps a problem description to the technique this course teaches for it.

Python Which technique fits which problem?

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

This is a simplification (real problems often mix several categories), but it captures the mindset: first understand the shape of the problem, then pick the algorithm family.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which description best fits classical (symbolic) AI?

    Classical AI works from a human-written description of the problem and applies algorithms such as search and logical inference.
  2. 2. What is an AI winter?

    There were at least two: in the 1970s and in the late 1980s to 1990s. Overpromising and brittle systems led to cuts in funding.
  3. 3. When is a classical solver usually a better choice than a neural network?

    Exact rules plus a need for provable correctness or explanations are the sweet spot for search, constraint solving and planning.
  4. 4. What did AlphaGo show about the relationship between the families of AI?

    AlphaGo used neural networks to guide and evaluate a Monte Carlo tree search, a hybrid of deep learning and classical search.

Summary

  • Classical AI solves problems with explicit models: search, constraint solving, logic, planning and probabilistic reasoning.
  • The three families (symbolic, statistical, deep learning) differ in where knowledge comes from; strong systems often combine them.
  • AI history alternates between optimism and AI winters; the lasting lesson is to be clear about what each method requires.
  • Classical methods shine when rules are known, guarantees and explanations matter, and data is scarce.
  • Every algorithm in this course runs live in Python in your browser.

Next, learn the vocabulary that all of AI shares in Agents and Environments.