An AI FSM Word Recogniser [$86]

More about An AI FSM Word Recogniser [$86]

Continuing from [$7F] i.e. the use of state-machines in pattern recognition (instead of neural nets) imagine we have a state-machine based parser where each state has 26 branches (or exits). This is very similar to the ::SHE+ILA:: parser, which has about 40 exits from each state. Now imagine we have a stream of upper-case ASCII alphabetic input. Each character uses 5 bits, and position register supplies a further 3 bits. We could thus decode plaintext words up to 8 characters long. (In practice our demo will be satisfied in recognising plaintext words of three characters, since this is the length of most Assembler mnemonics, e.g. TAD, DCA, AND, JMS, JMP, ISZ, OPR, IOT). The first level decodes the first character and branches 26 ways. The second level decodes the second character and has 26 states, each branching 26 ways. The third level has 26*26 states, and branches 26^3 ways (about 8,000), each exit representing one of the possible three-letter words. We train the FSM matrices by either the "weasel" algorithm or some sort of backpropagation algorithm, using a prose input of realistic three-letter words (like FOO, BAR, YOU, WOO, YAY, ETC) until only the first N outputs (where N is the number of words in the training sample) are produced. Larger FSMs could recognise OCR text, or bitmap images, but would need many more states. Also we should consider probabilities, as certain words are found together more frequently (e.g. FOO BAR, WOO YAY, MOO COW, etc.) Just a little experiment :) Can it write the works of Shakespeare, lol? Hey hey we're the Monkeys :)