I got 8th place. This was an interesting problem.
The task was to predict 8 strata locations for each well log. The data given are resistivity and gamma radiation per each depth. At first I tried to train lstm, but failed. Then I tried nn weighted by distance and it gave me decent provisional score ~750k. But I've not used well logs data at that point at all, the problem was I weighted not corrected strata locations. After researching a bunch of papers to correlate multiple well logs - found this paper "Incremental correlation of multiple well logs following geologically optimal neighbors" pretty decent. I've implemented correlating approach from there (in the order of euclidian distance) and after some tweaking it gave me 860k score. The problem I've not solved (there was no time and ideas) is how to choose good neighbors for correlation for given log. In my approach I just used max(8) nearest logs, corrected there strata locations, and weighted them by distance. The problem is that some near logs are not correlated well with given log which cause loss of score. I've tried ncc and several other quality metrics but no luck. The one quality metric is some kind of good is the distance(x,y) - where x - weighted gt strata locations, y - weighted correlated strata locations
The task was to predict 8 strata locations for each well log. The data given are resistivity and gamma radiation per each depth. At first I tried to train lstm, but failed. Then I tried nn weighted by distance and it gave me decent provisional score ~750k. But I've not used well logs data at that point at all, the problem was I weighted not corrected strata locations. After researching a bunch of papers to correlate multiple well logs - found this paper "Incremental correlation of multiple well logs following geologically optimal neighbors" pretty decent. I've implemented correlating approach from there (in the order of euclidian distance) and after some tweaking it gave me 860k score. The problem I've not solved (there was no time and ideas) is how to choose good neighbors for correlation for given log. In my approach I just used max(8) nearest logs, corrected there strata locations, and weighted them by distance. The problem is that some near logs are not correlated well with given log which cause loss of score. I've tried ncc and several other quality metrics but no luck. The one quality metric is some kind of good is the distance(x,y) - where x - weighted gt strata locations, y - weighted correlated strata locations
No comments:
Post a Comment