Chapter 10: Processes of Perception and Analysis

Section 9: Statistical Analysis

Markov processes

The networks in the main text can be viewed as representing finite automata (see page 957) with probabilities associated with transitions between nodes or states. Given a vector of probabilities to be in each state, the evolution of the system corresponds to multiplication by the matrix of probabilities for each transition. (Compare the calculation of properties of substitution systems on page 890.) Markov processes first arose in the early 1900s and have been widely studied since the 1950s. In their first uses as models it was typically assumed that each state transition could explicitly be observed. But by the 1980s hidden Markov models were being studied, in which only some of the states or transitions could be distinguished by outside observations. Practical applications were made in speech understanding and text compression. And in the late 1980s, building on work of mine from 1984 (described on page 276), James Crutchfield made a study of such models in which he defined the complexity of a model to be equal to -p Log[p] summed over all connections in the network. He argued that the best scientific model is one that minimizes this complexity—which with probabilities 0 and 1 is equivalent to minimizing the number of nodes in the network.

From Stephen Wolfram: A New Kind of Science [citation]