World Model

Here is the Blerpl learning algorithm’s view of the world:


Nothing too controversial here - hopefully we can agree that the world contains some causes, and that we receive information about the world from our senses. The structure of the world is what determines how those causes influence our sensor inputs. Blerpl’s job is to build a model of the world and discover that structure and those high-level causes, using only the sequences of information arriving at the inputs.

Each circle is a node, representing a cause. In Blerpl, causes are binary. That means that at any given moment, each node has a state, either zero or one. You might protest that the world contains causes with more than two states - Blerpl just models those as groups of binary causes. The choice of binary keeps the algorithm simple and general.

Nodes in0, in1, in2 and in3 are the sensor input nodes, like the input switches on the Predictatron, or the pixels in an image sensor. Of course there could be a lot more than four of them.

Nodes a, b, c, d and e are high-level independent causes in the world. They determine what is going on in the world. Because they are independent, the only basis we have for predicting them is the statistics of what they have done in the past. Independent nodes are drawn with a bold outline.

The box is a model of the structure of the world. It contains a network that describes how the independent causes influence our inputs. The network (not shown) is made up of nodes called dependent causes, because their states are determined by their parent nodes. The network may contain many layers of nodes, but ultimately every dependent cause’s state is determined by the independent causes.

You can think of the world model network as a box of and/or/not gates, flip-flops and so on. The assumption here is that the world can be modelled with boolean logic. And if signals turn up that can’t be predicted with logic, they are modelled as - guess what - independent causes.