4. Foundations of AI — Economics, Neuroscience, Psychology, and Linguistics
Source: AIMA 4th Ed, §1.2
Economics
Economics asked: How should agents make decisions to maximize payoff?
Key Contributions
| Figure / Concept | Contribution |
|---|---|
| Adam Smith (1776) | Agents pursuing individual self-interest → collective benefit; foreshadowed utility maximization |
| Jeremy Bentham / John Stuart Mill | Utilitarianism: actions should maximize total well-being (utility) |
| von Neumann & Morgenstern (1944) | Decision theory: how rational agents should choose under uncertainty using expected utility |
| Leonard Savage (1954) | Subjective expected utility (SEU): agents have subjective beliefs (probabilities) and preferences (utilities) → choose the action with highest expected utility |
| Markov Decision Processes (MDPs) | Mathematical formalism for sequential decision making under uncertainty — central to RL |
| John Nash (1950) | Nash equilibrium in game theory: no agent can unilaterally improve by changing strategy. Relevant for multi-agent AI. |
| Herbert Simon (satisficing) | In practice, agents don’t optimize perfectly — they “satisfice”: find a solution that is good enough. Rationality is bounded. |
Why It Matters for AI/RL
- Utility theory → the performance measure / reward function in RL
- Decision theory → the basis for rational agent behavior under uncertainty
- MDPs → the formal foundation for almost all of reinforcement learning
- Game theory → multi-agent RL, mechanism design, safety
Neuroscience
Neuroscience asked: How does the brain compute? What can we borrow?
Key Facts
| Topic | Key Finding |
|---|---|
| Neurons | ~100 billion neurons, each connected to ~1,000–100,000 others via synapses |
| Firing rate | Neurons fire at up to ~1,000 Hz; inter-neuron communication via electrochemical pulses |
| McCulloch & Pitts (1943) | First mathematical model of a neuron: binary threshold unit. Showed networks of such units could compute any computable function. |
| Hebb (1949) | Hebbian learning: “Neurons that fire together wire together” — synaptic strength increases when pre- and post-synaptic neurons fire simultaneously |
| Brain-machine interfaces | Modern work allows direct reading of motor intentions from neurons (e.g., paralyzed patients controlling computer cursors) |
Key Caveat
The brain is NOT a simple digital computer. Its architecture is massively parallel, fault-tolerant, and operates on analog signals. AI draws inspiration from it but is not a literal simulation of it.
Psychology
Psychology asked: How do humans and animals think and act?
Behaviorism vs. Cognitive Psychology
| School | Core Claim | Key Figure |
|---|---|---|
| Behaviorism (1913–1950s) | Only observable behavior matters; reject all talk of internal mental states | Watson, Skinner |
| Cognitive psychology (1960s–) | Internal representations and processes exist and explain behavior; the mind is an information processor | Craik (1943), Miller (1956) |
Craik’s Knowledge-Based Agent Framework (1943)
Kenneth Craik proposed that rational behavior can be explained by three steps: 1. The stimulus (percept) is translated into an internal representation 2. The representation is manipulated by cognitive processes (reasoning) 3. The result is retranslated into action
This is essentially the architecture of a rational agent in AIMA.
Miller’s “Magic Number 7” (1956)
George Miller showed that short-term memory can hold roughly 7 ± 2 items — one of the first rigorous results connecting psychology to information theory.
Control Theory
Control theory asked: How do devices use feedback from the environment to act optimally?
| Figure | Contribution |
|---|---|
| Norbert Wiener (1948) | Cybernetics: designed feedback controllers; connected feedback systems to purposeful behavior and intelligence |
| Optimal control (1950s–60s) | Minimize a cost function (e.g., fuel use, error) over time → precursor to the RL reward signal |
| Kalman filter (1960) | Optimal linear estimator for noisy systems — used in robotics and navigation |
Key difference from AI (historically): Control theory used continuous math (ODEs, optimization); early AI used symbolic logic. They are now converging — RL is the bridge.
Linguistics
Linguistics asked: What is the structure of language, and can machines understand it?
| Figure | Contribution |
|---|---|
| Chomsky (1957) | Formal grammars (context-free grammars) for syntax; showed behaviorism couldn’t explain language acquisition |
| Whorf hypothesis | Language shapes thought (controversial — largely rejected) |
| Computational linguistics | Formal study of language structure, parsing, generation, and understanding by computer |
| Knowledge representation | Language understanding requires world knowledge, not just syntax — drove knowledge representation research in AI |
Summary: What Each Field Gave AI
| Field | Core gift to AI |
|---|---|
| Philosophy | Justification that mind can be mechanical; logic as reasoning |
| Mathematics | Formal reasoning, computation limits, probability |
| Economics | Decision theory, utility, MDPs, game theory |
| Neuroscience | Neural network inspiration, Hebbian learning |
| Psychology | Agent architecture (percept → representation → action) |
| Control theory | Feedback, cost minimization, continuous optimization |
| Linguistics | Language structure, knowledge representation |