search

TheodoreEhrenborg / mercantilism

MIT License
0 stars 0 forks source link

Bayesian needs to be tested with many trials to see if it gives weird results #8

Closed TheodoreEhrenborg closed 5 years ago

TheodoreEhrenborg commented 5 years ago

So the versions of bayesian that use Monte Carlo integration do give weird results when there are hundreds of trials. I need to switch to symbolic integration.

TheodoreEhrenborg commented 5 years ago

The weird results come from the fact that of all the possible universes, only a very tiny fraction of them are even plausible considering the results. The computer spends most of its time searching the implausible universes. When it finds a plausible one, that universe has an outsize impact on the results. If symbolic integration does not work out, I can always try to limit the computer's search to universes near the most representative one.

TheodoreEhrenborg commented 5 years ago

I wonder if symbolic integration would work better with sympy.

TheodoreEhrenborg commented 5 years ago

I think bayesian.py is working now, although I do have to let trials = 10000 before it is accurate to a few decibals. Further comments will be in Issue #21.