Does the Kern On algorithm prioritize reducing further suggestions?

[Jeremy Tribby]

I seem to remember than Kern On makes suggestions based on frequency of pairs, as they occur in large amounts of analyzed text. I was wondering if the Kern On algorithm also considers the optimal pairs needed to result in Kern On having no more suggestions? In other words, is it possible that taking all (or many) of Kern On's suggestions can actually lead to more suggestions than if you were to kern other things? I only have this on my mind because lately it feels like Kern On is making more suggestions prior to stopping than it used to, but I suppose I could be imagining this — maybe Kern On is just getting better and judging my spacing more harshly :-)

[Tim Ahrens]

When Kern On determines the suggestions, the objective is to “fill the gaps” in the cloud of models. It should suggest pairs that have a shape combination which is not similar to any of the existing models, and therefore the engine (all models combined) currently has a “loose grip”, i.e. you will typically see many tickmarks because the possible span is large. This also implies that if the model is set then this will tighten the grip of the engine (including the new model) on the autpairs in general, in particular those that are similar to the new model. There are several other mechanisms at work when suggestions are generated, which are a bit difficult to explain here, but you can think of Kern On saying “I am unsure about this pair, please check,” and also “Having this as a model would help me be more certain about many autokerning pairs.”

Each suggestion candidate gets a score that reflects all the above, and also the frequency, as it seems more sensible to suggest a frequent pair rather than an infrequent one that is only slightly more unique in terms of shape combination.

Originally, Kern On was programmed to never stop suggesting pairs as I believed the users would stop adding models when they are satisfied with the overall result. I realised that some users would continue to follow suggestions forever so I implemented a simple threshold for the score, below which the suggestions are not be displayed and it sops at some point. This does not really mean Kern On considers the set of models to be “complete”, though, it is just a primitive threshold. To answer your original question: The suggestions that are shown to the user are, in a way, chosen so as to reach this point quicker (i.e. there are no more candidates above a certain threshold).

I know it would be nice if Kern On, at some point, simply said “Thanks, I have enough information, you are done,” but it’s not easy to define.

I have some ideas on how to streamline this whole process, maybe showing a whole text (synthesized like my Test Text Generator does) instead of presenting the suggestions one after the other. If this proof text looked right you could assume you have enough models.

[Jeremy Tribby]

Hi Tim, thank you for the detailed explanation, I understand the suggestions much better now