Artifact: activations tend to be lower when there are more propositions in a biasing analog.

[mars0i]

I think there's a general phenomenon, which I originally noticed in popco1, that if you use analogs to bias a distribution over runs, as I did with crime3, then having more propns in your biasing analog(s) will tend to pull the mean of mean activations down and toward the left in a scatterplot. This didn't matter much in crime3, because the number of propositions in each analog was either 12 or 13--i.e. similar numbers. But one saw the effect when using the beast and virus analogs together (the "both" panel in the 3-panel scatterplot I used in my Dispersal presentations). The red X, i.e. the mean, was below and left of 0,0, while effect of each of virus or beast individually, pushed the mean past 0 in the direction of effect.

I think I recall that I earlier decided that this was due to the kludge of making all analogy-net-derived negative links in the proposition network have a meta-weight relative to positive links, rather than making their relative effects a function of how many negative links there are. The problem, I think, is that if there are more propositions, then there would tend to be more competition, i.e. more negative links, so the absolute activation values tend to be pulled down. (In popco2 this is implemented by popco.nn.update/calc-propn-link-wt, which is indirectly called by update-propn-wts-from-analogy-activns, and which uses constants in +analogy-to-propn-pos-multiplier+, and +analogy-to-propn-neg-multiplier+ in popco.core.constants.)

In the past I just finessed this by having similar numbers of propositions in analogs. Is there a more general solution, though? Not necessarily a perfect one, but a good modeling compromise.

[mars0i]

A crude idea would be to adjust the ratio between negative and positive meta-weights (i.e. multipliers) +analogy-to-propn-pos-multiplier+, and +analogy-to-propn-neg-multiplier+ as a function of the size of the biasing analog.

Note, though, that popco2 doesn't really know anything about source and target analogs, and doesn't know that you're treating one side as the bias and then restricting communication to and measuring what happens in the other side. What popco2 knows is that make-analogy-net accepts two distinct sets of propositions, and that the order of parts of node names in the analogy net is a function of the other of those two sets.

I suppose another crude idea would be to allow the user to specify what the ratio is.

[mars0i]

For the specific task of comparing distributions in two dimensions due to different biases on other propositions (e.g. in my scatter plots), one solution might simply be to divide by the number of propositions, or use some normalizing ratio between them, when creating the scatter plots or creating the distributions that will be statistically tested. That's not a general solution, but it might be good enough. It does require extra explanation and justification when writing or speaking, but if I do it consistently and explain the motivation, it might be a reasonable solution.