Unit 2: Goals and Foundations of AI

Unit 2: Goals and Foundations of AI#

Session 3 · Aug 24 · HW2 assigned Aug 24, due Aug 30

Before you can evaluate an AI system, you need to know what it is trying to do. This unit builds the vocabulary you will use for the rest of the semester, and the vocabulary matters more than it might seem. Most public confusion about AI comes from mixing up categories: calling a chatbot “intelligent,” calling a classifier “biased,” calling a prediction a “decision.” Precision here pays off everywhere later.

Learning Objectives#

After completing this unit, you will be able to:

Distinguish the four goals AI has historically pursued and explain which ones dominate today.
Define the rational agent framework and use it to describe any AI system.
Classify a real-world problem into the correct task type (classification, regression, clustering, generation, and others).
Use core terminology, such as model, feature, label, training, and inference, correctly and precisely.

Part I — What Is AI Trying to Do?#

1.1 Four Historical Goals#

Russell and Norvig organize the field’s ambitions along two axes: whether the system should think or act, and whether the standard is human performance or rationality.

	Humanly	Rationally
Thinking	Cognitive modeling: replicate human reasoning	Laws of thought: reason correctly by logic
Acting	Turing Test: behave indistinguishably from a person	Rational agents: do the right thing given the goal

Modern AI overwhelmingly pursues the bottom-right quadrant: rational action. A spam filter is not trying to think like a person about email. It is trying to make the correct sorting decision as often as possible.

This is why “does the machine really understand?” is often the wrong question for practical purposes and simultaneously the right question for ethical ones. Hold both.

1.2 The Agent Framework#

Nearly any AI system can be described with four elements:

Environment → the world the system operates in
Sensors / Percepts → what it can observe
Actions / Actuators → what it can do
Performance measure → how we score whether it did well

Note

The performance measure is where most real-world AI failures originate. A system optimizes exactly what you measure, not what you meant. A recommendation engine measured on “watch time” will learn to maximize watch time. Future units return to this repeatedly.

1.3 Types of Environments#

Property	Easier	Harder
Observability	Fully observable (chess)	Partially observable (driving in fog)
Determinism	Deterministic	Stochastic
Agents	Single agent	Multi-agent, possibly adversarial
Time	Episodic (each case independent)	Sequential (actions have consequences)
Change	Static	Dynamic

Chess is fully observable, deterministic, and static, which is why it fell to computers first. Driving is none of those things, which is why it remains hard.

Part II — Core Terminology#

These terms below will appear in every remaining unit.

Term	Meaning	Example
Feature or Characteristic	An input variable the system observes	The symptoms a doctor notes
Label or Desired output	The correct answer for a given example	The confirmed diagnosis
Model	The learned mapping from features to output	The doctor’s accumulated judgment
Training	The process of fitting the model to data	Years of medical residency
Inference	Using the trained model on new input	Seeing a new patient today
Parameters	The internal numbers adjusted during training	What actually changed in the doctor’s head
Hyperparameters	The numbers of values that have to be defined to train or run the model	Model architecture
Generalization	Capability of the model to respond appropriately to inputs not previously seen	A new case the doctor has to deal with

A model is not a program someone wrote. It is a set of numbers or parameters that were discovered by a training process. Nobody chose them, and often nobody can fully explain them. That distinction drives the interpretability problems of some AI solutions.

2.1 Task Types#

Most AI problems reduce to a handful of types:

Classification: predict a category. Is this transaction fraudulent?
Regression: predict a number. What will this house sell for?
Clustering: group similar items with no labels given. What customer segments exist?
Ranking: order items by relevance. Which results go on page one?
Recommendation: match items likely to be of interest. What should this viewer watch next?
Generation: produce new content. Write a summary of this report.
Sequential decision-making: choose actions over time. Route this delivery fleet.

Correctly identifying the task type is one of the first steps in every AI project.

⚙️ Hands-On: Building a Rational Agent#

Here is a complete rational agent in about twenty lines. It lives in a small grid world, cannot see the whole map, and must find its way to a goal. Watch how the four elements (environment, sensors, actions, performance measure) appear explicitly in the code.

from collections import deque

# --- The ENVIRONMENT: 0 = open, 1 = wall ---------------------------
grid = [
    [0, 0, 0, 1, 0],
    [1, 1, 0, 1, 0],
    [0, 0, 0, 0, 0],
    [0, 1, 1, 1, 0],
    [0, 0, 0, 1, 0],
]
start, goal = (0, 0), (4, 4)

def neighbors(cell):
    """The agent's SENSORS: which adjacent squares are open?"""
    r, c = cell
    for dr, dc in [(-1, 0), (1, 0), (0, -1), (0, 1)]:   # its ACTIONS
        nr, nc = r + dr, c + dc
        if 0 <= nr < len(grid) and 0 <= nc < len(grid[0]) and grid[nr][nc] == 0:
            yield (nr, nc)

def find_path(start, goal):
    """Breadth-first search: guarantees the SHORTEST path."""
    queue = deque([[start]])
    visited = {start}
    while queue:
        path = queue.popleft()
        if path[-1] == goal:
            return path
        for nxt in neighbors(path[-1]):
            if nxt not in visited:
                visited.add(nxt)
                queue.append(path + [nxt])
    return None

path = find_path(start, goal)

# --- The PERFORMANCE MEASURE: number of steps taken ----------------
print(f"Path found in {len(path) - 1} steps:")
for r in range(len(grid)):
    row = ""
    for c in range(len(grid[0])):
        if (r, c) == start:   row += " S "
        elif (r, c) == goal:  row += " G "
        elif (r, c) in path:  row += " . "
        elif grid[r][c] == 1: row += " # "
        else:                 row += "   "
    print(row)

Try changing it:

Add a wall that blocks the path entirely. What does the agent do? Is failing gracefully a form of rational behavior?
This agent has no learning whatsoever, it searches fresh every time. Is it still “intelligent”? Defend your answer using the four goals from Part I.
Challenge: Alter the agent snippet such that random grids (variable sizes and walls) can be generated. Run your code a number of times and analyze the agent behavior.

Important

This agent is entirely symbolic, that is, the rules of movement were written by a human. It is a direct descendant of 1970s AI, and it works perfectly here because the environment is small, fully observable, deterministic, and static. Change any one of those and it breaks. That is the boundary line where machine learning becomes necessary.

💡 Example: Naming the Task Type#

For each scenario below, decide the task type before reading the answer.

Scenario	Task type
Flagging tumors in an X-ray	Classification
Estimating tomorrow’s electricity demand	Regression
Grouping students by study habits, no categories given	Clustering
Drafting a first-pass contract	Generation
Deciding which ad to show, learning from clicks over time	Sequential decision-making
Sorting job applicants by fit	Ranking (and an ethics problem — see Unit 10)

That last row is deliberate. Framing a human decision as a ranking task is itself a consequential choice, not a neutral technical one.

🧭 Reflection#

Pick an AI system you used this week. Write out its environment, sensors, actions, and performance measure. Now ask: is the performance measure it optimizes the same as the outcome you wanted? Where do they diverge?

📘 Further Reading#

Russell, S., & Norvig, P. (2022). Artificial Intelligence: A Modern Approach, 4th Ed., Ch. 2. Pearson.
Mitchell, M. (2020). Artificial Intelligence: A Guide for Thinking Humans, Ch. 1–3. Penguin Books.
Dendritic Institute (2025). AI Literacy Series — Module 2: AI and Its Many Branches.