Created: 26 Apr 2014 | Modified: 23 Jul 2020 | BibTeX Entry | RIS Citation |
Given a quick profiling, there aren’t obvious standouts, I’m simply doing a lot of operations many many times. But the graph library (networkx) is slow compared to others with a C/C++ backend, and I do graph operations tens of millions of times in a standard run. So that’s the obvious optimization, rewrite with a faster graph library.
I could do more static equilibrium analysis, but most of the interesting empirical hypotheses involve dynamics. In particular, four experiments suggest themselves after building some infrastructure for it:
Holding the size of the design space and the innovation rate constant, what happens to radius and other measures as the learning rate evolves (increases)? This is a better analysis of what happens with greater fidelity learning than comparing two static analyses with different parameters.
Holding learning rate and innovation rate constant, what happens when the design space evolves? I can implement this by configuring a design space with 16 trees, but only initializing the population from 4 of the trees (or whatever), but allowing individual innovation to access all 16. Over a long time frame, how does the fraction of design space filled evolve? If we keep track of the “level” of innovations – whole new trees, or leaves on an existing tree, what do the rates of major and minor technical innovations look like? Does breadth and radius evolve together over time?
In an evolving design space, what happens when we change the learning rate, holding the innovation rate constant?
In an evolving design space, what happens when we increase both fidelity, and the individual innovation rate as the design space increases in size?