In the examples we have given, the evaluation function is set in advance and the program forges ahead, searching nodes in the tree until it finds a suitable path to the solution. But, given a task such as solving a jigsaw puzzle or playing a game, a human subject will typically work step by step, inspecting the state of the problem and considering alternative strategies for proceeding toward the goal. Rather than evaluating all the permissible, or `legal', moves from the current state, a human expert will immediately eliminate most of them and may consider only one or two, reassessing the problem in the light of current knowledge. Indeed, outside of games and formal puzzles, it is not clear that the notion of legal moves is particularly helpful: a computer program that generates and tests legal moves in a search through a state-space may be operating in a far less intelligent, and certainly less efficient, way than a human being who thoughtfully `muddles through' towards a solution.
Some tasks will, in addition, involve the problem-solver in communication with an external source. A doctor diagnosing an illness, for example, will bring expert knowledge to bear not only on the patient's initial symptoms but also on information derived from tests, or elicited from the patient at a particular stage in the diagnosis. If, for instance, the patient has complained of breathlessness and a dull pain in the left arm, it would be appropriate for the doctor to ask whether the patient smokes.
Finally, a state-space is a description of the problem itself, rather than of the behaviour of the problem-solver; to this extent, it can be described independently in individuals' attempts to solve the problem. Maps of London, for example, antedate a particular person's efforts to find a route between two points in the city. If we consider the problem as it appears in the mind of the human problem-solver, then each new state exists only once the subject has generated it.