Adversarial Search and Games Questions and Answers

Answer: C

The basic algorithm assumes two players alternating, opposed utilities and a fully visible position, and each assumption relaxed produces a different variant: vectors for more players, chance nodes for randomness and belief states for hidden information. Listing the assumptions of a method is the quickest way to see which extensions exist and why each was needed.

Answer: D

Adding chance nodes removes the requirement that a move leads to a single known successor, which is what backgammon and card games need. Player count, observability and zero sum structure are relaxed by other variants. Identifying which single assumption an extension removes makes the family of algorithms a structured set rather than a list, and it predicts what each extension will cost.

Answer: A

Reasoning over sets or distributions of possible positions removes the assumption of full observability, at the cost of an exponentially larger effective state space. Determinism, opposition and turn order are relaxed by other variants. The pattern is consistent: each assumption removed buys generality and costs computation, and knowing the price in advance is what makes the choice deliberate.

Answer: B

Searching tens of plies with strong pruning, a good evaluation and prepared opening and endgame knowledge accumulates an advantage no human calculation can match. The tree is never enumerated, game databases help only in the opening, and the moves considered are the same legal moves a human sees. Attributing a result to the mechanism that actually produced it, rather than to a vaguer notion of superiority, is what makes the achievement instructive.

Answer: C

Combining a network evaluated policy and value with tree search raised playing strength sharply and the best balance between them is not settled. Programs have exceeded human strength since the nineteen nineties, the rules are unchanged, and chess is very far from solved. A field that continues after its headline milestone has been passed is usually one where the milestone turned out to be a poor summary of the underlying questions.

Answer: D

Chess was mastered by techniques that transferred poorly to perception and language, and go by different techniques again, so each success remained narrow. Computing power was necessary but not sufficient in either case, human expertise proved a mixed guide, and no generality followed. A sequence of milestones each achieved by a method suited to that game is exactly the pattern that the narrow and general distinction was drawn to describe.

Answer: A

A game supplies precise rules, an unarguable outcome and unlimited cheap practice, which makes progress measurable in a way most real tasks do not. The corresponding weakness is that the knowledge required is unlike that of messy real problems. Many games contain uncertainty, and transfer is exactly what does not follow. A testbed's virtues and its limitations are usually the same property seen from two directions.

Answer: B

Rules that are complete, states that are fully specified and outcomes that are unambiguous remove precisely the difficulties that dominate perception, language and physical action. That is why the transfer has been so limited. Difficulty, measurability and computation are not the objection. A criticism that identifies what a benchmark leaves out is more useful than one about how hard the benchmark is, since the omission is what limits what success there means.

Answer: C

Attention moved to translation, image recognition, dialogue and open ended tasks, where success is judged by human assessment rather than by a rule and the knowledge required is far broader. That shift brought its own difficulties of evaluation. Bigger games, theorem proving and numerical optimisation are all active but none carries the same visibility. A change of benchmark usually signals a change in what the field believes the hard part to be.

Answer: D

A translation or a summary may be good in several incompatible ways and bad in others, so no rule decides it and human judgement, with all its variability and cost, becomes the measure. Automatic proxies correlate imperfectly. Computation, recording and judge availability are practical matters. When the criterion of success is itself contested, progress becomes harder to establish, and that difficulty is inherited by everything built on the benchmark.