bkovitz / acclivation Goto Github PK

A Brief History -or- How All These Files Got In This Repository
---------------------------------------------------------------

In 2017, Ben Kovitz wrote some spreading-activation code in Clojure and
tried running an evolutionary simulation of knobs+graph organisms
(where knobs provide initial values to propagate by spreading activation
through the graph to determine a phenotype at some designated nodes in
the graph). It produced some visually spectacular acclivated virtual fitness
functions with "buttes", but it wasn't clear what the result was.

In 2018, Ben Kovitz and Marcela Poffald ran a bunch of experiments with
varying parameters on the spreading-activation algorithm, and some
"hill-climber" algorithms to try to measure how much acclivation occurs
and under what conditions. Experiments took all night, and acclivation
seemed to occur unpredictably. We still had to look at plots of the
virtual fitness functions to see if acclivation had occurred. Under
some parameters, it occurred seldom; tiny changes to the parameters would
make it occur rarely, and it wasn't at all clear why.

Dave Bender and Ben Kovitz rewrote the whole thing in C with some Python
support to make it easier to see plots of the various functions. The
simulations ran in seconds rather than hours, but it still wasn't clear
what makes acclivation likely vs. unlikely. We added more parameters and
tried more experiments but we did not get predictable results. Even under
more-favorable parameter settings, acclivation did not as often as we though
it should, given the simple idea for why it should work--and our old
observations of runs with now-forgotten parameter settings where acclivation
happened on almost every run.

In 2019, we tried again, inspired by a suggestion from Etienne Barnard to
try feed-forward (as opposed to spreading-activation) networks and to
allow larger populations. The feed-forward networks performed even worse,
but the new experiments led us to figure out that only a narrow range of
the S constant in the transfer function (see the ALIFE paper) works, and
why: only then do cyclic compositions of the transfer function create a highly
tunable set of ranges of expansion and contraction of the global fitness
function. The Clojure code in 2017 happened to have the S constant in this
range. It had been chosen for an unrelated spreading-activation experiment
in order to put attractors at -0.5 and +0.5.