The point of this exercise is to explore a connectionist version of selectionism, by training different networks on a 9-parity problem. Two independent variables are manipulated: the starting states of the weights, and the number of hidden units used in a network. Can connectionist networks be used to select a solution to this type of problem?
If the 9-parity problem turns out to be too much to ask your computer to deal with, do the exercises with the 7-parity problem instead. If you use this smaller network, then use 21 hidden units in Section 23.2.1 (instead of 27), use 14 hidden units in Section 23.2.2 (instead of 18), and use 7 hidden units in Section 23.2.3 (instead of 9). All of the other settings that are indicated for the 9-bit version (learning rates and such) should be OK.