The wire-move learning algorithm (Wuensche 1992a) moves a single wire at each cell position where there is a mismatch between the actual successor cell and the relevant cell in the target state. There is a choice of wires to select from the pseudo-neighbourhood, and a choice of new coupling positions, some of which will correct the mismatch. The aspiring pre-image order, the pseudo-neighbourhood order, and the coupling order, are all shuffled at random, and the first successful move is selected (if it exists).
Pre-existing (or just learnt) pre-images may be ``forgotten``, i.e. detached as pre-images of the target state. If ``just learnt'' states are forgotten when learning the next pre-image, the next wire/coupling alternative in the shuffled order will be tried. If none succeed the algorithm continues with the next aspiring pre-image. On completing the learning pass, the states learnt and not learnt are listed by their decimal equivalents (for n<=32 only), with the following prompt,
206,947,no (examples are Hamming distance 1 from 462)
l468,460,458,454,478,494,398,334,ok
re-learn by wire-moves -w, bit-flips -b: (if all ok, examples are Hamming distance 1 from 462)
or if all are ``ok``,
success - aspiring pre-images learnt
continue-ret:
``ok'' indicates learnt states, ``no`` not learnt. An option is offered to re-learn again by wire moves, or by bit-flips (which always succeeds). If ``w'' is selected the last ``no'' state will be learnt first, followed by the set of shuffled ``ok'' states, and the remaining shuffled ``no'' states. On completing the pass the resultant listing and prompts are repeated. Note that for a large system with a large list of aspiring pre-images, a wire-move learning pass may take an unpredictably long time.
To ``forget'' a pre-image just one appropriate wire move is required. Forgetting is generally much easier an quicker than learning. As fewer network changes are needed there will be fewer side effects. Note that in the forgetting procedure ``ok'' indicates states still attached to the target state, i.e. not forgotten, and ``no'' indicates states successfully forgotten, for example if all are ``no'' the following prompt is presented below the listed states,
success - aspiring pre-images forgotten
continue-ret:
If ``return'' is entered at the prompts above, the following prompt is presented,
more-m
Enter ``m'' to repeat the procedure from #27.5, otherwise the network as it stands is accepted.