Basin Saddle Discovery -- poor performance.

[Marchowes]

Basin Saddle Discovery for 0.6.0 is providing accurate results, but it takes hours to calculate because it searches the entire tree from every points perspective.

Possible new solution: Max recursion depth -- risky as it could sour results. ????? Optimizations? maybe the current model is poorly optimized and could use a little help?

Maybe add a feature that accurately exempts points we know are non basin saddles? something like: loop detection finds circuit, identifies basin saddle, takes all other points found in circuit and exempts them from further analysis. ?risky? -- worth testing at least. Analysis may still take a while, but should be shortened considerably. its the points that we know are non basin saddles that take so long to analyze because the entire tree needs to be searched.

[Marchowes]

My first run on a 1deg x 1 deg map took 16 hours -- well, thats no good. I was able to optimize several components and knock it down to 5 hours. Still unacceptable.

The solution to keep my current algorithm, and make it faster is to front load the remote linker checks, which take up 95% of the runtime (or more).

now, create dict {saddle.id : {"leg1": [linker1, linker2, linker3], "leg2":[linker1, linker2...]}} and read form that instead of calculating every time.

[Marchowes]

I've found a library "networkx" which can do what I need -- almost.

There is a weird problem tho with multiple conjoined cycles that requires me to re-run the whole cycle detection algorithm until no exceptions are found. This is a bit of a problem.

So, as a solution lets do the following:

if we find a loop disqualifier match exception we need to do the following.

1) when cycle detection is first run, we need to build an index of {saddle.id:[cycleidx_1, cycleidx_2, ....] 2) when we find an exception, we need to add the discovered cycle, as well as every single loop which contains that disqualified point. To a queue of points. 3) once the first round of analysis is done, take those points, and build a whole new Graph from scratch and run cycle_basis on that entire set. 4) rinse, repeat

what about a concept of disqualification candidates when 2 basins are of equal height?