Shannon's relay maze machine
In March 1951, at the Eighth Macy Conference, Claude Shannon showed a machine that finds its way through a 5 × 5 maze by trial and error, remembers the way and forgets it when the maze is changed. By his account it has about 75 relays, and the memory of each square is two relays holding one of four directions.
Why it matters
The transcript shows what «learning» could be in 1951: one memory cell per square and one rule, with no randomness and no stored program. This is an editorial assessment. The claim of the collection's editors that it was the first device of this level is not in the transcript itself and is not confirmed here.
What Shannon says. A maze of 5 × 5 = 25 squares with movable partitions; a «sensing finger» moved by two motors, east-west and north-south; the goal can be plugged into any square. The memory of a square is two relays; they hold the direction by which the finger left the square the last time. Exploration strategy: in a square the remembered direction is taken and rotated 90° counterclockwise, and rotated again if the finger hits a partition; the direction of entrance is skipped on the first rotation (without this, he says, solving took about three times as long). There is no randomness: «There is no random element present». Goal strategy: the direction by which the square was left last time is taken, so blind alleys and loops cancel themselves. Forgetting: if, after the goal is reached, the finger moved to another square does not arrive within 24 moves, the circuit reverts to the exploration strategy; the same counter breaks the «vicious circle» the machine can be led into when the partitions are moved so that the old solution runs in a loop (Gerard calls it a neurosis in the discussion, Shannon agrees). The exploration strategy, he says, solves any finite solvable maze. The relays, in his words, are «about seventy-five». What the machine is called. The transcript has neither the word Theseus nor the word «mouse»: it has «finger». The «magnetic mouse» Theseus, built in 1950, and a brain of «about 100 relays» are in the editors' introduction to the Collected Papers; that does not match the 75 relays Shannon gives in the talk, and the difference is not resolved here. What the record does not claim: how long or how many trials the learning took; the day of the talk (the source gives only the conference dates); that the machine was the first of its kind; where exactly it was built (no place is named in what was read); that the demonstration influenced any later work. The figures 25, 75, 24 and three times are the author's own description with no second source, hence the medium confidence.