| Speciation |
|
|
|
|
|
|
|
|
|
|
|
|
|
| • | Start by training
a single expert on all of the |
|||||||||||
| data. |
||||||||||||
| • | Then make two
copies of that expert, perturb the |
|||||||||||
| weights slightly,
and train the system of two |
||||||||||||
| experts and a
manager. |
||||||||||||
| – | This
allows each expert to be regularized by |
|||||||||||
| the
data that it does not eventually model. |
||||||||||||
| • | Then split each
expert again. |
|||||||||||
| – | You
have to decide how many times to split. |
|||||||||||