WIXLIB

ARTIFICIAL INTELLIGENCE FOR SOCIAL GOODbroadly to include all aspects of sustainable biological,economic, and social systems that support humanwell being. Here we focus primarily on the ecologicalcomponent, but the larger issues of social and economicsustainability must be considered as well.DataSeveral activities concern the measurement andcollection of data relevant to ecosystems.One approach is to develop and deploy sensornetworks. For example, the TAHMO (Trans-Africa HyrdoMeteorological Observatory; www.tahmo.org) project isdesigning and deploying a network of 20,000 weatherstations throughout sub-Saharan Africa (van de Giesen,2014). Several efforts are deploying camera traps tocollect image data or microphone systems to collectbioacoustic data. Still other projects employ unmannedaerial vehicles to obtain video imagery for trackingelephants and other large animals. AI algorithms can beapplied to optimize the locations of these sensors andtraps in order to gather the most valuable informationShort-term Applications and ChallengesCurrent research and applications in AI for sustainabilitycan be organized in terms of data, modeling, decisionmaking, and monitoring. The goal is to manageecosystems with policies that are based on the highquality data and science.Figure 2: Image-Based Ecological Information System4

goal is to understand which climate variables affectspecies habitat carrying capacity.at the lowest cost. Once the data are collected, other AIalgorithms can be applied to identify species and tracktheir locations.A second important type of model seeks to characterizethe spatio-temporal dynamics of species. Such modelscan predict migration, dispersal, reproduction, andmortality of species. Developing such dynamicalmodels is critical to developing policies that can helpendangered species thrive and control the spread ofinvasive species. One recent example is the BirdCastbird migration model under that combines eBird andweather radar to test hypotheses about the behaviorof migrating birds (birdcast.info).A second approach to data collection is to engage citizenvolunteers. One of the oldest citizen science projects iseBird (www.ebird.org), in which bird watchers uploadchecklists of the birds they have seen at a particulartime and place. The Image-Based Ecological InformationSystem (IBEIS; www.ibeis.org) project analyzes animalphotos scraped from internet sources such as Flickrand Facebook and applies computer vision and activelearning methods to detect the animals, identify thespecies, and even identify individual animals (“Bob thegiraffe”) (see Fig. 2). Their AI techniques can identifyunique animals as long as they have stripes, wrinkles,or other unique textures.Policy OptimizationOnce we have models of species distribution, behavior,and habitat requirements, we can begin to designand optimize policies for successful management ofspecies and ecosystems. This requires articulating ourpolicy goals and objectives. Virtually every ecosystemmanagement problem combines an ecological model withan economic model of the economic costs and benefitsof various policy outcomes.A third approach employs technically trained people(e.g., government and corporate scientists) to collectdata. One example of this is the freshwater streamsurveys conducted by the EMAP project of the html/), and similar efforts by forest resource companiessuch as Weyerhauser about-our-forests/uswest/). These groups collect samples of freshwatermacroinvertebrates that live in streams. These insectsmust then be examined by humans to identify the genusand species of each. AI computer vision methods arenow being applied to accelerate this process.One example is the design of a schedule for purchasinghabitat parcels to support the spatial expansion of theRed Cockaded Woodpecker (Sheldon, 2010; Sheldon, 2015).An optimal policy takes into account uncertainty in thespread of the species and the availability of land parcelswhile seeking to link up existing patches of reservedhabitat. Algorithms for computing this policy combineideas from network cascade analysis (maximizing spreadin social networks) with techniques from AI planning andMonte Carlo optimization.ModelsAfter data are collected, the data can be analyzed byapplying techniques from data mining, statistics, andmachine learning to discover trends and fit models. Forspecies data, most efforts begin by counting individualsin order to produce estimates of population size andmaps of the spatial distribution of species. For thispurpose, it is particularly valuable to identify individualsand take into account multiple detections of thesame individual across time and space. These modelscan support some inferences about species habitatrequirements. In light of climate change, an importantA second example considers the temporary needsof migrating birds. Instead of making a permanentpurchase of land, the Nature Conservancy is applyingdetailed bird migration models developed by the CornellLab of Ornithology to rent rice fields in California(Axelson, 2014). The farmers who own those fields agreeto flood them at the right time to support migratingwaterfowl in the pacific flyway. The exact timing variesfrom year to year based on the predictive migration5

ARTIFICIAL INTELLIGENCE FOR SOCIAL GOODmodels developed using multi-scale machine learningtechniques (cite STEM).game theoretic models, this work assumes thatthe adversary behavior is governed by modelsderived from past adversary behavior. The use ofsecurity resource allocation using game theoryhas been explored in urban contexts, particularlyin counter-terrorism settings in the past (Tambe,2011). The current work builds on this past workwhile also significantly enhancing it with machinelearning based predictions of adversary behaviors.A third example of policy development confronts theissue of long-term planning in the face of climatechange and sea level rise. One approach, known asadaptive management, explicitly considers the need toupdate and revise models based on new data that willbecome available in the future. There is considerableuncertainty about the timing and degree of sea levelrise. Nicol, et al., (2015) formalize the problem ofplanning in the presence of this uncertainty. Theirgoal is to conserve appropriate coastal habitat formigrating birds under the risk of sea level rise. Lowlying land will become inundated, and migrating birdswill be threatened unless additional habitat furtherinland is available for them. As always, there are tightbudgetary constraints on the amount of land thatcan be purchased. AI algorithms for solving PartiallyObservable Markov Decision Problems are able to solvethese difficult constrained optimization problems(Pineau, 2003).Opportunities and Challenges in theMedium TermWith few exceptions, most work in ecosystemmanagement and conservation focuses on a smallnumber of species in particular regions. A majorchallenge for the medium term is to develop methodsthat can collect and model data encompassing a broadrange of species at continental scales. This will requireintegrating many different data sources (e.g., for birds,fish, plants, and insects) collected by many differentmethods (e.g., stationary sensors, earth orbit satellites,citizen scientists, sensors worn by animals, and so on).There are many research issues in data management anddata integration that must be addressed. For example,the most common approach to data integration is toassimilate all data to a fixed spatial and temporal scaleby smoothing fine-scale data and interpolating coarsescale data. This process introduces distortions into thedata. We need methods for integrating and modelingdata at multiple scales that can retain the resolutionand uncertainty associated with each data source.Case Study: Monitoring and EnforcementWhen policies are put into effect, there is oftena need for law enforcement to ensure theirsuccessful execution and for monitoring to detecterrors in the data and models that require reoptimizing the policy. For example, elephantpoachers in Africa routinely enter national parksand other bioreserves to hunt elephants illegallyfor the ivory tusks. The PAWS project (Fang, 2016;Nguyen, 2016; Yang, 2014) applies AI algorithms topredict poaching attacks and optimize the patrolroutes of game wardens in order to maximizetheir deterrent effect while minimizing costs.PAWS relies on two key underlying systems: (i)Predicting poacher behavior from past poachingdata: This system builds poacher behavior modelsusing machine learning algorithms (Nguyen, 2016);(ii) Game theoretic security resource allocation:Using these learned poacher behavior models, ituses game theoretic algorithms to prescribe newpatrolling strategies; thus, rather than assuming thestandard rational adversary response in standardA related shortcoming of current modeling efforts isthat they generally assume stationary (steady-state)climate, land use, and species behavior whereas thereal systems are experiencing climate change, rapideconomic development, and continuing evolution,dispersal, and natural selection of species. Modelingtechniques and supporting data are needed that cantake into account these drivers of change and the manyuncertainties associated with them.As the scale of questions grow, it is no longer possibleto focus only on the biological components of a6

system. Instead, one must take a “systems of systemsapproach” and incorporate models of social, cultural,and economic activity. For example, when choosinga site for a new dam, we must consider not only theimpact on native and invasive species in the riverineecosystem, but also the benefits for farming, thepotential inundation of important cultural and religioussites, and changes in sediment transport that mayaffect the distribution of pollutants and contaminatedsoils. Current AI technologies cannot currently operateat this scale and level of complexity.A third challenging aspect of sustainability work arisesdue to lack of technical infrastructure: poor networking,little access to high-performance computing resources,and lack of local personnel with sufficient educationand training. We must develop algorithms that can runlocally on small computers (or telephones) that onlyhave intermittent access to large cloud computingresources. We must take into account the possibilitythat human actors may fail to adhere to designatedpolicies. Finally, we must develop creative methods ofestablishing metrics for assessing the effectiveness ofdata collection and policy execution to compensate forthe lack of historical data.One trend that will enable broader and morecomprehensive modeling of ecosystems is thecontinuing improvement of sensors: reduction in size,power requirements, and cost. These improvementswill support and drive the demand for better modelsthat can support the development of higher-qualitypolicies. However, cheaper sensors can be less reliable,so research is needed on methods for automaticallydetecting and removing bad data and broken sensors.This is a theme that is also common to other areas suchas healthcare and public policy, as we discuss below,and poses a major challenge for many AI-powereddecision-support systems.A fourth challenge is finding business models that supportlong-term data collection and policy enforcement efforts.Many current projects rely on the enthusiasm of citizenscientists, the generosity of private donors, or grantsfrom funding agencies. None of these is likely to providesteady, long-term support. One possibility is to developbusiness models that generate continuing revenuestreams. For example, the TAHMO project seeks to sellits weather data to insurance companies, commoditiestraders, and other businesses that rely on high qualityweather data and forecasts.A second set of data challenges concerns the biasesand quality of data, particularly crowd-sourced data.Birders choose where they go bird watching; touristsand tour operators choose where people take picturesof wildlife. Even the data collected by game wardensis biased by the need to maintain unpredictability.New incentive mechanisms are needed to encouragevolunteers to collect less biased data. Examples ofmechanisms that are showing some success includethe Great Zebra Challenge and the eBird Global BigDay. New algorithms are needed to incorporate datacollection goals into the PAWS enforcement games.And methods for explicitly modeling the data collectionprocess (“measurement models”) must be improved. Amajor analytical challenge is that when measurementmodels are incorporated into machine learning, thevariables of fundamental interest are no longer directlyobserved. This raises questions about the identifiabilityand semantics of the inferred values of those variables.Long Term ProspectsSustainability is concerned with the long-term health ofecosystems and human societies. As we contemplate thecreation and deployment of policies over the long term,we must confront the fact that the long-term behaviorof ecological, economic, and social systems is radicallyuncertain. We can be very confident that our currentmodels are missing critical variables and importantinteractions. How can artificial intelligence methods dealwith the uncertainty of these “unknown unknowns”?One important strategy is to plan for the “learningprocess”. When a new policy is put into place, we mustalso develop and deploy an instrumentation plan tocollect data on a broad range of variables. We mustincorporate “precautionary monitoring”, in which wemonitor not only the variables that we expect to changeas a result of the policy, but also a wide range of variables7

ARTIFICIAL INTELLIGENCE FOR SOCIAL GOODthat could allow us to detect unexpected side effectsand unmodeled phenomena. We must plan to iterativelyextend our models to incorporate these phenomena andre-optimize the policies.The Surgical Critical Care Initiative (SC2i), aDepartment of Defense funded research program, hasdeployed two clinical decision support tools (CDSTs) torealize the promise of precision medicine for criticalcare (Belard, 2016). The invasive fungal infectionCDST was deployed in 2014 to assist military providerswith treatment decisions both near point of injuryand at definitive treatment centers. Trauma-relatedinvasive fungal infections are well recognized for theirdevastating impacts on patients in both military andcivilian populations. In addition to substantial morbidityresulting from recurrent wound necrosis (e.g., greaternumber of surgical procedures, amputations, anddelayed wound closure), the disease is also associatedwith high mortality rates (Tribble and Rodriguez, 2014;Warkentien, 2012; Lewandowski, 2016; Rodriguez, 2014).Finally, when formulating and optimizing managementpolicies, we should adopt risk-sensitive methods.Standard practice in solving economic models is tominimize the expected costs and maximize the expectedbenefits of the policy. But if a policy has substantialdownside risk (e.g., species extinction, economiccatastrophe), then we should apply AI methods thatfind robust policies that control these downside risks.This is an active area of research (see, e.g., Chow, 2015),and much more work is needed to understand howwe can ensure that our models are robust to both theknown unknowns (as in traditional risk managementmethods) and the unknown unknowns.The massive-transfusion protocol (MTP) CDST iscurrently being assessed under a two-year clinical trialat Emory-Grady, one of the two SC2i civilian hospitals.This CDST uses evidence-based predictive analyticsto help physicians identify which patients genuinelyrequire a massive transfusion, thereby reducingcomplications associated with over-transfusion orthe needless expenditure of blood products (Maciel,2015; McDaniel, 2014; McDaniel, 2014; O’Keeffe, 2008;Dente, 2009). The SC2i is also planning a clinical trialaround its WounDX, a CDST that predicts the timing oftraumatic wound closure. Once validated, this tool hasthe potential to substantially improve outcomes (by asmuch as 68%) and reduce resource utilization ( 3.4Bannual cost-savings) both nationally and within theMilitary Health System (Forsberg, 2015).HealthSuccess StoriesCurrent methods for gathering population-scaledata about public health through surveys of medicalproviders or the public are expensive, time consuming,and biased towards patients who are already engagedin the medical system. Social media analytics isemerging as an alternative or complementary approachfor instantly measuring the nation’s health at largescale and with little or no cost. Natural languageprocessing can accurately identify social media poststhat are self-reports of disease systems, even for rareconditions. The nEmesis system, for example, helpshealth departments identify restaurants that are thesource of food-borne illness (Sadilek, 2016). nEmesisfinds all the Twitter posts for a city that are sent byrestaurant patrons, and then checking if any of thepatrons tweet about the symptoms of foodborne illnessover the next 72 hours. When this happens, healthdepartment officials are alerted of the fact, so that theycan schedule inspection of the restaurant. nEmesissignificantly improved the effectiveness of inspectionsin Las Vegas and the Center for Disease Control isfunding the expansion of nEmesis nationwide.Case Study: Making ‘Meaningful Use’meaningfulSepsis is the 11th leading cause of death in theUS – seven hundred fifty thousand patientsdevelop severe sepsis and septic shock in theUnited States each year. More than half of themare admitted to an intensive care unit (ICU),accounting for 10% of all ICU admissions, and 20%to 30% of hospital deaths. Yet others experiencesepsis due to hospital acquired infections8

(HAIs) in the medical units. Several studies havedemonstrated that morbidity, mortality, andlength of stay are decreased when severe sepsisand septic shock are identified and treated early;Kumar et al. 2006 show that every hour delay intreatment is associated with a 7-8% increase inmortality. Screening t

Artificial Intelligence (AI) is currently seeing major media interest, significant interest from federal agencies, and interest from society in general. From its origins in the 1950s, to early optimistic predictions of its founders, to some recent negative views put forth by the media, AI