http://tiny.cc/xomo

xomo: one short file that runs COCOMO-II (effort), COQUALMO (defects) and Boehm's risk tables (schedule/cost risk) as a Monte-Carlo over four classic NASA/JPL case studies (flight, ground, osp, osp2). A project is a box of rating ranges; sample inside it many times and read effort, defects and risk as distributions, not single guesses. The model is pure stdlib; --learn grows a min-variance decision tree over the draws using nuff.

# install and test
git clone http://tiny.cc/xomo && cd xomo
python3 -B xomo.py             # medians (p25 p50 p75) per study
python3 -B xomo.py --learn     # decision tree: what drives good outcomes
python3 -B xomo.py --checks    # self-tests

Sections: NAME | SYNOPSIS | OPTIONS | DATA | TESTS | OUTPUT | LEARN | CALIBRATION | SEE ALSO | LICENSE | AUTHOR

Files: xomo.py | Makefile | pyproject.toml

NAME

xomo - Monte-Carlo COCOMO-II + COQUALMO (effort, defects, risk)

SYNOPSIS

python3 -B xomo.py [-s N] [-n N] [-l N] [-p STUDY] [--learn] [--checks] [-h]

OPTIONS

-s --seed     random seed                   seed=1
-n --n        samples/study, or --learn rows  n=1000
-l --leaf     min rows per tree leaf         leaf=3
-p --project  study to --learn               project=osp
   --learn    grow a decision tree (needs nuff)
   --checks   run model self-tests
-h --help     show the docstring

DATA

A case study is a dict of COCOMO driver ranges. Any driver left unset spans its full legal range. Drivers fall in four families:

family   members                              role
------   -----------------------------------  ----------------------
sf       prec flex resl team pmat             scale factors (exponent)
em(+)    rely data cplx ruse docu time stor   effort up as rating up
         pvol
em(-)    acap pcap pcon aexp plex ltex tool   effort down as rating up
         site sced
dr       aa etat pr                           defect removers

Each draw: pick an integer rating in every driver's box, then

effort  = a * KLOC^(b + 0.01*sum SF) * prod EM        (COCOMO-II)
defects = sum over {reqs,design,code} of
          introduced(SF,EM) * (1 - removed(dr))        (COQUALMO)
risk    = 100 * (fired Boehm risk-table cells) / 216

TESTS

--checks runs every test_*:

test_positive                 effort/defects/risk finite, in range
test_osp_riskier_than_osp2    maturing osp->osp2 cuts risk
test_flight_bigger_than_ground tighter flight s/w costs more
test_seed_repeats             same seed -> same draw
test_learn_tree               nuff tree finds at least one split

OUTPUT

# 1000 samples/study, seed=1
                          effort               defects               risk
  study      p25     p50     p75     p25    p50    p75    p25   p50   p75
 flight      921    1971    3672    1606   3508   6939      4     6     7
 ground      558    1144    2123    1508   3471   6846      2     3     5
    osp     1743    2240    3014    3589   5511   8161     12    16    20
   osp2      237     307     386     131    205    298      3     3     4

Note osp -> osp2: process maturity (prec, pmat up) drops median effort ~7x, defects ~27x, and risk from 16 to 3 -- the "orders of magnitude" effect, shown reproducibly.

LEARN

--learn turns N random draws of one project into a table (driver ratings -> effort/defects/risk) and grows nuff's min-variance decision tree over it. The root split = the single factor that most separates good from bad outcomes; follow branches to the + leaf for the rating combo that minimises the blended goals.

python3 -B xomo.py --learn -p osp2 -n 200 -l 20

   d2h    n  Effort-  Defects-  Risk-  tree
  0.51  200   324.21    228.95   3.48              <- baseline (all draws)
  0.45  148   317.55    224.21   3.12  Sced > 2
  0.37   79   260.86    194.78   3.11  |  Kloc <= 102
+ 0.31   36   231.56    170.25      3  |  |  Ltex > 3
  ...

Needs nuff: pip install nuff, or clone the sibling gist so $DOOT/nuff is importable. -n sets rows, -l the min leaf size.

CALIBRATION

Each draw guesses a fresh internal calibration (effort a,b, the EM/SF slopes, COQUALMO m-ranges). So 1000 runs marginalize over the space of plausible calibrations: a conclusion that survives is stable no matter which calibration is "true" -- letting you decide without local tuning data. This is the core XOMO idea (Menzies & Richardson, 2005). Tweak the ranges in the POLICY block to widen or narrow that space.

SEE ALSO

konfig    http://tiny.cc/konfig   shared Makefile/boilerplate
nuff      http://tiny.cc/nuff     tiny stdlib python tricks
paper     Menzies & Richardson, "XOMO: Understanding Development
          Options for Autonomy", 20th Intl Forum on COCOMO, 2005.
          http://timmenzies.net/pdf/05xomo.pdf

LICENSE

MIT. https://choosealicense.com/licenses/mit/

AUTHOR

Tim Menzies, timm@ieee.org

built by gistsite