You are on page 1of 8

Recent

 Advances  in  AI  for  Computational  Sustainability  


AI  and  Sustainability  Department,  IEEE  Intelligent  Systems  

Douglas  H.  Fisher    


Vanderbilt  University  
 
COPYRIGHT  owned  by  IEEE;  this  accepted  version  posted  with  permission;  Fisher,  
D.,  "Computing  and  AI  for  a  Sustainable  Future,"  IEEE  Intelligent  Systems,  vol.  27,  no.  
4,  July/Aug  2012.  
 

Though  research  at  the  intersection  of  AI  and  sustainability  has  been  going  on  for  
many  years,  with  papers  such  as  Todorovski,  et  al  [1]  and  Phillips,  et  al  [2]  being  
some  of  the  more  recent  examples  in  this  long  history,  growth  in  the  area  
accelerated  coincident  with  the  founding  of  the  International  Conferences  on  
Computational  Sustainability.    

The  inaugural  conference,  CompSust09  (http://www.computational-­‐


sustainability.org/compsust09),  highlighted  the  relevance  of  AI  to  problems  
concerning  sustainability,  with  areas  of  machine  learning  and  optimization  taking  
center  stage,  and  computer  vision  receiving  substantive  attention  too.  Of  
sustainability  areas,  particular  attention  was  given  to  the  natural  environment,  
including  biodiversity,  natural  resource  protection,  and  Earth  system  monitoring.  In  
CompSust10  (http://www.computational-­‐sustainability.org/compsust10),  the  
topics  addressed  opened  up  considerably,  and  included  energy-­‐efficient  computer  
architectures  and  algorithms;  social  computing,  crowdsourcing,  and  collective  
intelligence;  as  well  as  other  areas  of  AI,  such  as  computational  game  theory.  
Sustainability  areas  beyond  those  directly  concerned  with  the  natural  environment  
also  strongly  emerged,  most  notably  energy,  smart  grids,  electricity  markets,  
transportation,  and  materials.  These  new  themes  had  been  presaged  in  earlier  work,  
such  as  white  papers  of  the  CRA/CCC  (http://www.cra.org/ccc/whitepapers.php).  

While  there  was  no  CompSust  conference  in  2011,  chairs  of  the  earlier  conferences,  
Carla  Gomes  and  Brian  Williams,  launched  a  special  track  at  the  Twenty-­‐Fifth  
Conference  on  Artificial  Intelligence  (AAAI-­‐2011)  on  Computational  Sustainability  
and  AI,  which  has  carried  over  to  AAAI-­‐2012  as  well.  This  was  a  significant  
development,  because  it  incorporated  sustainability  into  a  flagship  conference  of  AI.  
In  fact,  a  paper  in  the  special  track  [3]  received  one  of  the  conference’s  two  
outstanding  paper  awards.      

Because  this  IEEE  IS  issue  is  temporally  coincident  with  the  AAAI-­‐2012  and  the  
Third  International  Conference  on  Computational  Sustainability  
(http://www.computational-­‐sustainability.org/compsust12/),  it  seems  fitting  to  
survey  the  papers  in  the  earlier  2011  conference.  Importantly,  these  papers  reflect  a  
representative  sample  of  the  research  underway,  with  attention  to  sustainability  
aspects  of  built,  social,  and  natural  environments;  using  AI  strategies  for  learning,  
planning  and  scheduling,  and  optimization  generally.  An  overview  also  invites  
synthesis,  with  AI  concepts  and  methods  being  a  framework  for  reasoning  about  
sustainability  challenges.    

 
Figure  1:  Papers  of  the  AAAI-­‐2011  special  track  on  Computational  Sustainability  are  organized  
vertically  by  a  main  AI  thrust  and  horizontally  by  a  main  sustainability  theme.  Each  paper  is  
indicated  by  an  abbreviated  title,  and  by  the  number  of  the  paper  as  it  is  listed  in  the  References  
section.  The  figure  is  intended  to  give  a  quick,  global  picture  of  the  papers,  recognizing  that  a  2D  
representation  oversimplifies.  

AI  and  the  Natural  Environment  


Several  papers  in  AAAI-­‐11  advanced  optimization  methodologies  for  purposes  of  
resource  management.  Lai,  et  al  [4]  addresses  the  problem  of  designing  corridors  
that  link  up  existing  protected  reserves  of  land,  typically  so  that  an  endangered  
species  can  mix  genetically  within  the  protected,  contiguous  region.  The  paper  
observes  that  corridor  design  can  be  framed  as  a  Steiner  tree  problem  of  connecting  
two  or  more  specified  nodes  in  a  graph  (i.e.,  the  existing  reserves)  by  a  least-­‐cost  
network  of  nodes  (i.e.,  land  parcels),  where  cost  might  be  the  purchase  price  of  
acquired  parcels.  In  [4],  the  problem  of  building  a  least-­‐cost  corridor  that  supports  
multiple  species  is  considered.  The  authors  term  this  the  Steiner  Multigraph  
Problem,  and  for  corridor  design  it  acknowledges  that  not  all  land  parcels  may  be  
suitable  to  all  species  (e.g.,  one  species  may  find  a  steep  mountainside  impassable,  
whereas  another  may  not).  
 
There  are  interesting  variants  of  corridor  design.  The  suitability  of  land  can  be  
generalized,  for  example,  from  binary  (suitable  or  unsuitable)  to  ordinal  or  
continuous,  and  a  highest-­‐utility  subgraph  (i.e.,  corridor)  that  does  not  violate  a  
fixed  budget  is  desirable  [5].  This  variant  better  approximates  what  a  real  wildlife  
manager  may  be  confronted  with  –  not  how  do  I  find  the  least  cost  solution,  but  
rather,  how  do  I  find  the  highest  utility  solution  subject  to  my  fixed  (monetary)  
budget?  In  another  AAAI  paper,  Golovin  et  al  [3]  address  the  general  problem  of  
obtaining  parcels  of  land  that  maximize  utility  (i.e.,  survival  probability  of  one  or  
more  species),  subject  to  a  budget  constraint.  While  these  authors  are  concerned  
with  the  acquisition  of  reserves,  and  not  corridors  linking  existing  reserves,  they  
address  the  challenge  of  dealing  with  a  budget  that  unfolds  piecemeal,  each  year  a  
new  budget  that  (typically)  only  partially  satisfies  a  long-­‐term  vision.  Moreover,  
land  parcel  costs  and  utilities  can  change,  for  example,  as  residential  lots  encroach  
on  formally  isolated  habitats.  This  dynamic  resource  allocation  [3]  problem  reflects  
more  closely  still  what  a  wildlife  manager  is  confronted  with  –  changing  budgets,  
utilities,  and  costs.  
 
Reserve  and  corridor  design  are  clearly  related  and  complementary,  so  it  would  not  
be  surprising  to  see  dynamic  variants  on  corridor  design  emerge.  Moreover,  these  
efforts  have  developed  relatively  efficient  algorithms  that  do  a  good  job  of  
approximating  the  best  solutions  possible  in  their  respective  tasks.    More  generally,  
“probably  approximately  optimal”  algorithms,  which  find  solutions  that  are  
sufficiently  close  to  optimal  most  of  the  time,  are  often  desirable  if  they  are  also  
computationally  cheaper.  Often  the  “efficiently  probably  approximately”  
characteristics  stem  not  from  formal  algorithmic  guarantees,  but  are  an  empirical  
manifestation  of  patterns  and  constraints  found  in  real-­‐world  problem  structure  [4].    
 
Dynamic  planning  begs  more  sophisticated  models  of  changing  environments  so  
that  planning  algorithms  can  project  into  the  future.  In  [3]  a  focus  on  modeling  
interactions  between  species  and  environment,  for  example,  informs  projections  
that  a  species  will  occupy  the  parcel  or  not.  Models,  if  available  for  corridor  design  
[4],  could  be  used  to  estimate  the  probability  a  species  would  stay  within  a  corridor  
or  wander  into  unprotected  regions.  Phillips  and  Elith  [6]  and  Hutchinson,  et  al  [7]  
discuss  machine  learning  for  predicting  occupancy  of  parcels  by  species,  and  both  
can  be  folded  into  larger  scale  wildlife  and  resource  modeling  and  planning  systems.    
 
The  importance  of  modeling  agents,  such  as  species  (to  include  humans)  as  a  
function  of  context,  could  be  taken  further  by  evaluating  designs  against  full-­‐blown  
agent-­‐based  simulations  [8].  In  fact,  this  is  essentially  the  motivation  behind  
Crowley  and  Poole’s  [9]  use  of  expert-­‐created,  high-­‐fidelity  simulations  of  forest  
dynamics  to  learn  timber  harvesting  policies  that  are  sustainable  over  decades.  
Policy  learning  strategies  are  also  the  focus  of  Petrik  and  Zilberstein  [10],  who  
develop  novel  approaches  that  are  applied  to  find  best  policies  for  water  reservoir  
release  (e.g.,  for  generating  hydro-­‐power).  Though  this  later  project  relies  on  
historical  data  to  learn  policies,  it  too  could  benefit  from  simulations  of  river  and  
reservoir  dynamics.  Policy  learning  (e.g.,  where  and  how  much  to  harvest,  when  and  
how  much  water  to  release)  is  suited  to  these  latter  projects  because  human  
interventions  are  intended  to  be  transitory  (e.g.,  forests  regenerate  and  reservoirs  
refill),  and  indeed  the  intervention  is  part  of  a  sustainable  cycle.  In  contrast,  reserve  
and  corridor  design  are  intended  as  interventions  that  are  more  or  less  permanent.    
 
AI  and  the  Built  Environment  
A  number  of  papers  were  concerned  with  applications  in  human  made  
environments,  and  we  focus  here  on  buildings  and  transportation.    

El-­‐Din  Mady,  et  al  [11]  describe  an  HVAC  controller  to  lower  energy  usage  and  lower  
expected  occupant  discomfort.  HVAC  actions  are  informed  a  Markov  model  of  
building  occupancy,  with  random  elements  modeling  uncertainty  on  adherence  to  a  
standard  work  schedule.  There  is  also  impressive  modeling  of  the  room  
environment,  including  the  heat  dynamics  of  walls,  water  radiators,  windows  and  
the  like  –  all  of  this  is  required  for  high-­‐fidelity  simulations  of  changing  building  
conditions.  Occupant  discomfort  is  measured  as  the  squared  difference  between  the  
actual  and  a  referent  temperature,  weighted  by  a  probability  of  occupancy.    

Experimental  simulations  indicate  the  promise  and  limitations  of  this  model-­‐
dominant  approach.  Input  from  occupant  sensors  and  counting  the  number  of  
occupants  is  planned  as  future  work,  and  indeed  the  addition  of  actual  observations  
would  be  a  natural  and  helpful  input  to  the  controller.  We  might  even  imagine,  with  
attention  to  privacy  concerns,  that  computer  vision  activity  recognition  could  
identify  individuals  and  customize  conditions  to  their  stated  or  learned  preferences.  

In  contrast  to  the  model-­‐dominant  approach,  Kolter  and  Ferreira  [12]  evaluate  
various  (regression)  approaches  that  learn  from  data  to  predict  a  building’s  total  
energy  usage.  The  project  collated  data  from  GIS  databases  on  building  locations  
and  layouts,  tax  assessor  data  on  building  characteristics,  and  data  on  monthly  
electricity  and  gas  usage,  then  compared  the  accuracy  of  various  regression  
methods.  With  good  predictors  in  hand,  a  building’s  actual  energy  usage  can  be  
compared  to  predicted  usage  for  buildings  of  like  characteristics  –  this  can  be  useful  
information  in  the  hands  of  a  building  manager  who  can  then  investigate  and  adjust  
energy  use.  Presumably,  those  buildings  doing  worse  than  predicted  can  benefit  
from  occupant  modeling  [11]  and/or  even  greater  customized  control.  
 
Three  papers  were  concerned  with  transportation,  specifically  with  personal  
vehicles  over  short  distances  by  road.  Two  of  the  papers  are  concerned  with  using  
AI  search  methods  to  find  optimal  routes,  where  optimality  is  defined  differently  in  
each  case.  Sachenbacher  et  al  [13]  are  concerned  with  finding  routes  for  electric  
vehicles  that  minimize  net  energy  loss,  where  some  segments  (edges)  in  a  route  may  
have  negative  costs  because  these  segments  have  properties  that  serve  to  recharge  
the  vehicle’s  battery.  The  system  relies  on  a  model  of  energy  dynamics  in  order  to  
estimate  route  segment  costs,  and  these  estimates  are  made  dynamically,  based  on  
changing  factors  such  as  payload  weight  and  AC  usage.    

Apple,  et  al  [14]  also  plan  routes  to  minimize  expected  time  of  transit.  These  routes  
are  based  on  modeling  assumptions,  as  well  as  data  from  smart  phones.  To  assist  in  
this  planning,  the  system  models  traffic  light  control,  enabling  it  to  compute  
probability  distributions  over  time  spent  at  lights  by  a  vehicle  conditioned  on  
various  factors.  Another  source  of  uncertainty  that  is  dealt  with  in  [14]  is  the  
location  and  route  taken  by  a  vehicle,  given  GPS  data  on  vehicle  location.    

Along  different  lines,  Au  et  al  [15]  extend  earlier  work  in  multi-­‐agent  systems  that  
move  smart  vehicles  through  intersections  with  no  reliance  on  traffic  lights.  Rather,  
smart  vehicles  request  passage  and  are  guided  through  as  requests  are  honored.  
Simulations  show  that  the  new  intersection  management  protocol  reduces  delays  at  
intersections  in  the  face  of  unbalanced  traffic  flows  from  inlet  roads.    In  future  
contexts  of  ample  (smart)  resources,  we  might  expect  the  technology  of  [15]  to  
reduce  uncertainties  that  are  currently  modeled  in  the  routing  approaches  of  [13]  
and  [14],  because  the  vehicle  itself  intervenes  in  the  process.  
 
Inference,  Data,  Control,  and  Design  
Papers  of  AAAI-­‐12’s  computational  sustainability  track  illustrate  different  
commitments  to  model-­‐based  inference  and  data-­‐supported  inference,  ranging  from  
model-­‐dominant  at  one  end  (e.g.,  [11])  to  data-­‐dominant  at  the  other  (e.g.,  [12]).  
Undoubtedly,  as  projects  move  out  of  simulation  and  into  the  physical  world,  they  
will  increasingly  use  data.  Li  and  Williams’  [16],  who  describe  planning  for  an  
autonomous  underwater  vehicle  (AUV)  for  purposes  of  ocean  monitoring  illustrates  
data  collection  through  robotic  exploration,  which  could  be  adapted  as  part  of  
aquatic  reserve  design.  Quinn,  et  al  [17]  describe  another  data-­‐collection  technology  
that  informs  occupancy  assessment  using  camera-­‐enabled  mobile  devices  to  identify  
crop  disease.    While  they  are  directly  concerned  with  ‘occupancy’  of  disease,  the  
technology  is  promising  for  purposes  of  other  kinds  of  occupancy  assessment.  While  
data  comes  with  its  own  stochastic  elements,  it  can  also  reduce  uncertainty  for  
problems  such  as  occupancy  detection  of  various  forms.  
 
Controlled  interventions  can  further  lower  uncertainty  (or  raise  it!).  Many  papers  
propose  interventions  into  natural  and  built  settings,  but  also  social  settings.  
Though  we  have  not  focused  on  this  latter  case,  two  papers  explored  market  
strategies  for  smart  grid  operations,  one  for  controlling  the  use  of  “home”  and  other  
micro  storage  of  energy  (Voice,  et  al  [18])  and  one  exploring  implications  of  
different  small-­‐scale  energy  broker  strategies  on  supplier  balances  and  other  
balances  (Reddy  and  Veloso  [19]).    
 
Apropos  control,  Santhanam,  et  al  [20],  investigates  model-­‐checking  strategies  for  
evaluating  the  success  of  interventions  that  are  intended  to  stop  “infections”  on  
networks,  be  they  social,  geographic,  physical,  or  economic.  The  emphasis  on  these  
and  the  great  majority  of  papers  on  evaluation  with  respect  to  interventions  is  
noteworthy.  If  anything,  investigations  need  to  expand  the  context  in  which  
evaluation  occurs,  to  ferret  out  what  may  be  otherwise  unanticipated  consequences  
of  technological  interventions  by  humans.  
 
Finally,  interventions  are  often  forwarded  to  make  the  best  out  of  poorly  designed  
environments.  For  example,  attempts  to  improve  the  efficiency  of  travel  by  private  
automobiles  that  use  fossil  fuels  are  a  good  example  –  these  efforts  are  important,  
but  we  can  go  further  and  design  systems,  ranging  from  products  to  communities,  to  
be  efficient  and  to  invite  efficiency  in  the  manner  in  which  they  are  used.  Many  of  
the  papers  are  overtly  design  papers  (e.g.,  corridor  design).  Additionally,  there  are  
two  papers  that  focus  on  more  conventional  forms  of  design.  Patnaik,  et  al  [21]  
induce  lifecycle  assessment  trees  for  products  from  “high  level”  product  descriptions.  
Such  trees  can  be  used  in  the  design  of  products,  reasoning  about  the  environmental  
impacts  of  component  design  choices  for  example,  favoring  environmentally  sound  
choices.  Hoenigman,  et  al  [22]  addresses  the  design  of  residential  landscapes  that  
require  less  watering;  this  is  more  than  using  native  plants,  but  is  cast  as  an  
optimization  problem,  getting  into  issues  of  plant  arrangement  in  ways  that  heartier  
plants  protect  more  sun-­‐sensitive  plants.  While  this  might  be  considered  a  non-­‐
traditional  design  problem,  it  nicely  illustrates  how  environmental  design  upfront  
can  mitigate  the  need  for  playing  environmental  catch-­‐up  later.  
 
Final  Remarks  
This  article  intended,  at  a  minimum,  to  give  the  briefest  glimpse  of  the  papers  of  the  
2011  AAAI  Computational  Sustainability  track,  so  that  interested  readers  could  
follow-­‐up.  Collectively,  however,  the  contributions  include  representations  of  much  
that  we  might  hope  to  see  in  an  environmentally  minded  cognitive  agent,  with  
competencies  in  sensing/observation,  knowledge-­‐based  reasoning,  decision-­‐
making,  and  actuation.  The  creation  of  such  agents  could  be  an  exciting  and  
important  ambition,  with  attention  paid  to  agent  communication  as  well.  In  the  near  
term,  however,  AI  will  offer  important,  albeit  specialized  cognitive  tools  for  decision  
makers.  It  is  critical  that  these  tools  are  adopted,  adapted  and  used  by  humans.  
Thus,  it  is  vital  that  as  part  of  our  evaluation  methodologies,  we  move  beyond  
questions  concerning  technical  optimality,  and  that  we  work  with  social,  behavioral  
and  economic  scientists  to  understand  the  social  contexts  in  which  our  tools  are  
used  (or  not)!  

References  
1.  L.  Todorovski,  et  al.,  “Using  equation  discovery  to  revise  an  Earth  ecosystem  
model  of  carbon  net  production,”  Ecological  Modelling,  no.  170,  2003,  pp.  141-­‐154.  
2.  S.  J.  Phillips,  et  al.,  “Maximum  entropy  modeling  of  species  geographic  
distributions,”  Ecological  Modelling,  no.  190,  2006,  pp.  231–259  

3.  D.  Golovin,  et  al.,  “Dynamic  Resource  Allocation  in  Conservation  Planning,”  AAAI  
Conference  on  Artificial  Intelligence,  AAAI,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3617.    
 
4.  K.  Lai,  et  al.,  “The  Steiner  Multigraph  Problem:  Wildlife  Corridor  Design  for  
Multiple  Species,”  AAAI  Conference  on  Artificial  Intelligence,  AAAI,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3768.    

5.  J.  Conrad,  et  al.,  “Incorporating  Economic  and  Ecological  Information  into  the  
Optimal  Design  of  Wildlife  Corridors,”  Computing  and  Information  Science  Technical  
Reports,  http://hdl.handle.net/1813/17053.  Cornell  University,  Ithaca,  NY,  2010.  

6.  S.  Phillips  and  J.  Elith,  “Logistic  Methods  for  Resource  Selection  Functions  and  
Presence-­‐Only  Species  Distribution  Models,”  AAAI  Conference  on  Artificial  
Intelligence,  AAAI,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3657.  
 
7.  R.  Hutchinson,  et  al.,  “Incorporating  Boosted  Regression  Trees  into  Ecological  
Latent  Variable  Models,”  AAAI  Conference  on  Artificial  Intelligence,  AAAI,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3711.  
 
8.  E.  Bonabeau,  “Agent-­‐Based  Modeling:  Methods  and  Techniques  for  Simulating  
Human  Systems”  Proc.  Nat’l  Academy  of  Sciences,  vol.  99,  suppl.  3,  2002,  pp.  7280–
7287;  www.pnas.org/content/99/suppl.3/7280.full.  
 
9.  M.  Crowley  and  D.  Poole,  “Policy  Gradient  Planning  for  Environmental  Decision  
Making  with  Existing  Simulators,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3631.  
 
10.  M.  Petrik  and  S.  Zilberstein,  “Linear  Dynamic  Programs  for  Resource  
Management,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3608.    
 
11.  A.E.  Mady,  et  al.,  “Stochastic  Model  Predictive  Controller  for  the  Integration  of  
Building  Use  and  Temperature  Regulation,”  AAAI  Conference  on  Artificial  
Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3719.  

12.  J.  Kolter  and  J.  Ferreira,  “A  Large-­‐Scale  Study  on  Predicting  and  Contextualizing  
Building  Energy  Usage,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3759.  
13.  M.  Sachenbacher,  et  al.,  “Efficient  Energy-­‐Optimal  Routing  for  Electric  Vehicles,”  
AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3735.    

14.  J.  Apple,  et  al.,  “Green  Driver:  AI  in  a  Microcosm,”  AAAI  Conference  on  Artificial  
Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3648.  

15.  T.  Au,  et  al.,  “Enforcing  Liveness  in  Autonomous  Traffic  Management,”  AAAI  
Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3762.    

16.  H.  Li  and  B.  Williams,  “Hybrid  Planning  with  Temporally  Extended  Goals  for  
Sustainable  Ocean  Observing,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3667.    

17.  J.  Quinn,  et  al.,  “Modeling  and  Monitoring  Crop  Disease  in  Developing  Countries,”  
AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3777.  
 
18.  T.  Voice,  et  al.,  “Decentralised  Control  of  Micro-­‐Storage  in  the  Smart  Grid,”  AAAI  
Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3632.    
 
19.  P.  Reddy  and  M.  M.  Veloso,  “Learned  Behaviors  of  Multiple  Autonomous  Agents  
in  Smart  Grid  Markets,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3773.    
 
20.  G.  R.  Santhanam,  et  al.,  “Verifying  Intervention  Policies  to  Counter  Infection  
Propagation  over  Networks:  A  Model  Checking  Approach,”  AAAI  Conference  on  
Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3742  
 
21.  N.  Sundaravaradan,  et  al.,  “Discovering  Life  Cycle  Assessment  Trees  from  Impact  
Factor  Databases,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3751.    
 
22.  R.  Hoenigman,  et  al.,  “Water  Conservation  Through  Facilitation  on  Residential  
Landscapes,”  AAAI  Conference  on  Artificial  Intelligence,  2011,  
http://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/view/3763  
 

You might also like