[Development] Objective measures

[vietly23]

Hi,

I'm going through the redesign document, and there is no mention of scoring functions/objective measures. Should we add an abstract class? I'm picturing the decision process as follows:

boolean AlignerClass::shouldChange() {
    Evaluator scoringFunction = this.getEvaluator(); // Generic scoring class
    Score newScore = scoringFunction.incrementalChange(source, oldTarget, newTarget);
    Score incScore = newScore - oldScore;
    float probability = calculateProbability(score, currentTemperature);
    return (incScore >= 0 or randomReal(gen) <= badProbability);
}

Obviously, the terminology can change, but what should we call the generalized class?

[red4711]

Yes, this would be a great idea. Pasha and I didn't get that far into the design. As far as SANA 2.0 is concerned, we should flesh out our skeleton design before we do any logic implementation.

As for the naming, you could call it something like ObjectiveFunction.

[waynebhayes]

Viet, Thanks! Why not just use "measure" as we do in SANA 1? Then there are two types of measures: those that are objectives (ie., those that are capable of guiding the alignment---these need to have an incremental evaluation operator), and those that are measures that we only compute after-the-fact, either because we don't care to use them to guide the alignment, or because we can't figure out how to compute them incremetnally. LCCS is one of the latter: it's a half-decent measure, but it would be hard to make in incrementally evaluated.... and in any case it's probably not worth using it as an objective, because it's absolutely trivial, at least on paper, to create an alignment with an LCCS of 1 if we're willing to sacrifice other objectives: just swap edges until the the largest common connected subgraph is the entire smaller graph. This would be hard to implement the accounting, but would result in alignments that are otherwise lousy anyway. Thus LCCS is not suited to be an *objective*, but it's a decent after-the-fact measure to evaluate an existing alignment. So it's not an objcetive even though it's a measure.

…

-- Wayne Hayes, Ph.D. Director, UCI-SDSU Joint Ph.D. Program in Computational Science (UCI side) Associate Professor, Computer Science, University of California, Irvine President Trump is a self-absorbed, self-serving, habitual liar. The former two are opinion; the latter is fact. Examples (1) http://foreignpolicy.com/2017/08/18/trumps-lies-about-james-comey-keep-unraveling . (2) https://www.theguardian.com/us-news/trump-administration

On Thu, May 24, 2018 at 11:38 AM, Linh Tam Chu ***@***.***> wrote: Yes, this would be a great idea. Pasha and I didn't get that far into the design. As far as SANA 2.0 is concerned, we should flesh out our skeleton design before we do any logic implementation. As for the naming, you could call it something like ObjectiveFunction. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#10 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ATUkjDa1w2Diyi_ikuvV3A1TJajoJsYjks5t1v4XgaJpZM4ULo26> .