Numbo Part 2

In the second half of the section on Numbo by Daniel Defays, he talks in detail about how Numbo works in theory followed by a description of some sample runs. As expected, Numbo works very similarly to Jumbo with some small differences. As previously mentioned, it features a Pnet as a sort of knowledge base and factors in randomness more so than Jumbo. Like Jumbo, it features a "cytoplasm" which works very similar for the most part to Numbo. The biggest difference between the two are the backtracking when a block is judged unattractive. Jumbo has a very simple way of backtracking up a pathway one level at a time, whereas Numbo is more volatile in how it removes bricks or destroys blocks altogether, like the 77 block in the first trace run of Numbo.

Overall, I feel that Numbo does a good job of simulating how a human would work one of these problems, or at least as good as it may get. As Defay mentions, Numbo does have some shortcomings in some simple problems, like the final puzzle he gives, but it is mostly due to the fact that you can only load the Pnet with so many proximity targets and facts. It is simply not possible to have a program be as knowledgeable as a capable human when it comes to its "knowledge base". Also as Defay suggests, it is very difficult, if not impossible, to fully describe how a human would solve one of these problems. Whether it is subconscious or just unnoticed, there will always be some steps left out. With that said, the randomness, Pnet, and other facets of the process of how Numbo runs make it as close as it can get to the human mind, at least for now.