Textbook of AI

Micro-worlds was the strategy, dominant at the MIT AI Laboratory in the 1960s and 1970s, of attacking artificial intelligence by working within small, fully-specified domains where every relevant entity, relation, and action could be enumerated and hand-coded. The bet was that competence built in micro-worlds would compose into competence in the broader world, a bet that turned out to be largely wrong, but which produced, along the way, much of the conceptual vocabulary still used in AI today.

The famous micro-worlds

• The blocks world, coloured blocks on a table; operators pick up, put down, stack. Made famous by Terry Winograd's SHRDLU (1971), which conducted natural-language dialogues about a simulated blocks world ("Find a block which is taller than the one you are holding and put it into the box"). SHRDLU integrated parsing, semantic interpretation, world modelling, and planning with what seemed, at the time, formidable competence.
• STUDENT (Daniel Bobrow, 1964), a micro-world for natural-language algebra word problems. STUDENT could parse and solve problems like "If the number of customers Tom gets is twice the square of 20% of the number of advertisements he runs, and the number of advertisements is 45, what is the number of customers Tom gets?"
• SAINT (James Slagle, 1961), symbolic integration of indefinite integrals at first-year-undergraduate level.
• Gelernter's Geometry Theorem Prover (1959), proofs of Euclidean geometry theorems with diagrammatic heuristics.
• Sussman's HACKER and SOPHIE (Brown, Burton & de Kleer), micro-worlds for simple electronic circuits, the latter an early intelligent tutoring system.
• MACSYMA, a more substantial symbolic-mathematics environment that grew out of this tradition.

The bet, and why it largely failed

The intellectual bet was articulated explicitly by Marvin Minsky and Seymour Papert: techniques perfected in micro-worlds would compose into competence on the broader world as the worlds grew. The bet largely failed. Each micro-world hid an enormous amount of world-specific knowledge, SHRDLU's apparent linguistic sophistication was inseparable from the closed ontology of its blocks; STUDENT's parser worked because algebra word problems use a tightly constrained sub-grammar of English. The techniques rarely generalised when the domain expanded.

The recognition of this failure to scale was one of the factors precipitating the first AI winter in the mid-1970s and contributed to the rise, in the late 1970s and 1980s, of expert systems, which embraced rather than denied the need for explicit, voluminous domain knowledge, encoded as production rules.

Quiet renaissance

The micro-worlds tradition had a quiet renaissance in the late 2010s, but as a methodology for benchmarking specific cognitive capabilities of large neural networks rather than as an engineering programme. Tasks such as:

• bAbI (Weston et al., 2015), 20 reasoning tasks in a tiny synthetic world.
• CLEVR (Johnson et al., 2017), visual question answering on rendered scenes of geometric shapes.
• ARC (Chollet, 2019), abstraction and reasoning corpus, deliberately small and synthetic.
• Procgen, MiniGrid, BabyAI, RL micro-worlds for sample-efficiency research.

are essentially micro-worlds for testing reasoning, visual question answering, abstraction, and generalisation. The shift in framing is telling: where 1970s micro-worlds were ends, modern micro-worlds are probes, instruments for diagnosing the capabilities of systems trained on much broader data.