Wednesday, February 20, 2013

TSE paper outline

1- The ICSE'13 result: All algorithms perform almost the same at 2 or 3 objectives. IBEA beats all by a large margin when optimizing 4 or 5 objectives.
2- The CMSBSE'13 result: Lower rates are better. This is a brief section, IBEA vs NSGAII vs SPEA2, E-Shop only, high rates vs low rates. It can be extended to other algorithms and feature models, but then it's dwarfed by the tree mutation results. 
3- The comparison of algorithms is run for 20 feature models, with effect size calculations... Confirms superiority of IBEA. Lower rates are used here to give all algorithms their best chance to perform.
4- Tree Mutation: By removing random crossover and mutation and deploying tree mutation [at a low rate 0.01], IBEA performs better (converges faster) than 1 & 2 above. 
4-a) The improvement is 3-orders-of-magnitude when comparing E-Shop with ICSE'13 setup vs Tree Mutation. In fact, we now have good results at 2 sec and very good at 10 sec, thus interactive responses are possible. [Comparison: With ICSE'13 IBEA, We achieve HV=0.28 & 52% correct solutions after 3 hours. With Tree Mutation IBEA, We achieve HV=0.25 & 99% correct solutions after 8 seconds. The 0.28 figure is inflated by incorrect solutions that achieve supreme objective values... 3hours/8sec = 1350].
4-b) The improvement can also be demonstrated on all 20 feature models, only the numbers are less than 3-orders-of-magnitude... [This needs to be repeated because of the bug we've just discovered].
5- Tree Mutation does not help other algorithms much... This can be demonstrated on all 20 models.  [This needs to be repeated because of the bug we've just discovered]. Should this be included? If we don't include it here, is it worthy of a whole new paper?
5-a) Comparison of 6 algorithms with and without Tree Mutation. found at of Methods Feb-Mar 2013.pdf

No comments:

Post a Comment