This article presents PAWS, a game-theoretic applicationdeployed in Southeast Asia for optimizing foot patrolsto combat poaching. In this article, we report on thesignificant evolution of PAWS from a proposed decision aidintroduced in 2014 to a regularly deployed application. Weoutline key technical advances that lead to PAWS’s regulardeployment: (1) incorporating complex topographic fea-tures, for example, ridgelines, in generating patrol routes; (2)handling uncertainties in species distribution (game-theoret-ic payoffs); (3) ensuring scalability for patrolling large-scaleconservation areas with fine-grained guidance; and (4) han-dling complex patrol scheduling constraints.

Poaching is a serious threat to wildlife conservation andcan lead to the extinction of species and destruction ofecosystems. For example, poaching is considered a majordriver (Chapron et al. 2008) of why tigers are now found inless than 7 percent of their historical range (Sanderson et al.2006), with three out of nine tiger subspecies already extinct(IUCN 2015). As a result, efforts have been made by lawenforcement agencies in many countries to protect endan-gered animals from poaching. The most direct and com-monly used approach is conducting foot patrols. However,given their limited human resources and the vast area inneed of protection, improving the efficiency of patrolsremains a major challenge.

Articles

SPRING 2017 23Copyright © 2017, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602

PAWS — A Deployed Game-Theoretic Application

to Combat Poaching

Fei Fang, Thanh H. Nguyen, Robert Pickles, Wai Y. Lam, Gopalasamy R. Clements, Bo An, Amandeep Singh, Brian C. Schwedock, Milind Tambe, Andrew Lemieux

n Poaching is considered a majordriver for the population drop ofkey species such as tigers, ele-phants, and rhinos, which can bedetrimental to whole ecosystems.While conducting foot patrols isthe most commonly usedapproach in many countries toprevent poaching, such patrolsoften do not make the best use ofthe limited patrolling resources.

Game theory has become a well-established para-digm for addressing complex resource allocation andpatrolling problems in security and sustainabilitydomains. Models and algorithms have been proposedand studied extensively in the past decade, formingthe general area of security games (Tambe 2011). Fur-thermore, several security-game-based decision sup-port systems have previously been successfullydeployed in protecting critical infrastructure such asairports, ports, and metro trains (Pita et al. 2008;Shieh et al. 2012; Yin et al. 2012). Inspired by the suc-cess of these deployments, researchers have begunapplying game theory to generating effective patrolstrategies in green security domains such as protect-ing wildlife (Yang et al. 2014; Fang, Stone, and Tambe2015), preventing overfishing (Haskell et al. 2014,Qian et al. 2014), and illegal logging (Johnson, Fang,and Tambe 2012).

Among these prior works, a novel emerging appli-cation called PAWS (protection assistant for wildlifesecurity) (Yang et al. 2014) was introduced as a game-theoretic decision aid to optimize the use of humanpatrol resources to combat poaching. PAWS was thefirst of a new wave of proposed applications in thesubarea now called green security games (Fang, Stone,and Tambe 2015; Kar et al. 2015). Specifically, PAWSsolves a repeated Stackelberg security game, wherethe patrollers (defenders) conduct randomizedpatrols against poachers (attackers) while balancingthe priorities of different locations with different ani-mal densities. Despite its promise, the initial PAWSeffort did not test the concept in the field.

This article reports on PAWS’s significant evolutionover the last two years from a proposed decision aidto a regularly deployed application. We report on theinnovations made in PAWS and lessons learned fromthe first tests in Uganda in spring 2014, through itscontinued evolution since then, to current deploy-ments in Southeast Asia and plans for future world-wide deployment. In this process, we have workedclosely with two nongovernment organizations (Pan-thera and Rimba) and incorporated extensive feed-back from professional patrolling teams. Indeed, thefirst tests revealed key shortcomings in PAWS’s initialalgorithms and assumptions (we will henceforth referto the initial version of PAWS as PAWS-Initial, and tothe version after our enhancement as PAWS). First, amajor limitation was that PAWS-Initial ignored topo-graphic information. Second, PAWS-Initial assumedanimal density and relevant problem features at dif-ferent locations to be known, ignoring the uncer-tainty. Third, PAWS-Initial could not scale to providedetailed patrol routes in large conservation areas.Finally, PAWS-Initial failed to consider patrol-sched-uling constraints.

In this article, we outline novel research advancesthat remedy the aforementioned limitations, makingit possible to deploy PAWS on a regular basis. First,we incorporate elevation information and land fea-

tures and use a novel hierarchical modeling approachto building a virtual street map of the conservationarea. This virtual street map helps scale-up while pro-viding fine-grained guidance and is an innovationthat would be useful in many other domains requir-ing patrolling of large areas. Essentially, the streetmap connects the whole conservation area througheasy-to-follow route segments such as ridgelines,streams, and river banks. The rationale for this comesfrom the fact that animals, poachers, and patrollersall use these features while moving. To address thesecond and third limitations, we build on the streetmap concept with a novel algorithm that uniquelysynthesizes two threads of prior work in the securitygames literature; specifically, the new PAWS algo-rithm handles payoff uncertainty using the conceptof minimax regret (Nguyen et al. 2015), while simul-taneously ensuring scalability — using our streetmaps — through the cutting plane framework (Yanget al. 2013). Finally, we incorporate into PAWS theability to address constraints such as patrol time lim-its and starting and ending at the base camp. In thefinal part of the article, we provide detailed informa-tion about the regular deployment of PAWS.

Background and Related WorkCriminologists have worked on the problem of com-bating poaching, from policy design to illegal tradeprevention (Lemieux 2014). Geographic informationsystems (GIS) experts (Hamisi 2008) and wildlifemanagement staff (Wato, Wahungu, and Okello2006) have carefully studied the identification ofpoaching hotspots. In recent years, software toolssuch as SMART2 and MIST (Stokes 2010) have beendeveloped to help conservation managers record dataand analyze patrols retrospectively. We work on acomplementary problem of optimizing the patrolplanning of limited security staff in conservationareas.

In optimizing security resource allocation, previ-ous work on Stackelberg security games (SSGs) hasled to many successfully deployed applications forthe security of airports, ports, and flights (Pita et al.2008; Fang, Jiang, and Tambe 2013). Based on theearly work on SSGs, recent work has focused on greensecurity games (Kar et al. 2015), providing conceptu-al advances in integrating learning and planning(Fang, Stone, and Tambe 2015) and the first applica-tion to wildlife security, PAWS-Initial. PAWS-Initial(Yang et al. 2014) models the interaction between thepatroller (defender) and the poacher (attacker) whoplaces snares in the conservation area (see figure 1) asa basic green security game, that is, a repeated SSG,where every few months, poaching data is analyzed,and a new SSG is set up enabling improved patrollingstrategies. The deployed version of PAWS adopts thisframework.

We provide a brief review of SSGs, using PAWS as a

Articles

24 AI MAGAZINE

Articles

SPRING 2017 25

key example. In SSGs, the defender protects T targetsfrom an adversary by optimally allocating a set of Rresources (R < T) (Pita et al. 2008). In PAWS, thedefender discretizes the conservation area into a grid,where each grid cell is viewed as a target for poach-ers, to be protected by a set of patrollers. The defend-er’s pure strategy is an assignment of the resources totargets. The defender can choose a mixed strategy,which is a probability distribution over pure strate-gies. The defender strategy can be compactly repre-sented as a coverage vector c = 〈ci〉 where ci is the cov-erage probability, that is, the probability that adefender resource is assigned to be at target i(Korzhyk, Conitzer, and Parr 2010). The adversaryobserves the defender’s mixed strategy through sur-veillance and then attacks a target. An attack couldrefer to the poacher, a snare, or some other aspectfacilitating poaching (for example, poaching camp).Each target is associated with payoff values that indi-cate the reward and penalty for the players. If theadversary attacks target i, and i is protected by thedefender, the defender gets reward Udr,i and the adver-sary receives penalty Uap,i. Conversely, if not protect-ed, the defender gets penalty Udp,i and the adversaryreceives reward Uar,i. U

ar,i is usually decided by animal

density — higher animal density implies higher pay-offs. Given a defender strategy c and the penalty andreward values, we can calculate the players’ expectedutilities Uai and U

di when target i is attacked accord-

ingly.In SSGs, the adversary’s behavior model decides his

response to the defender’s mixed strategy. Past workhas often assumed that the adversary is perfectlyrational, choosing a single target with the highestexpected utility (Pita et al. 2008). PAWS is the firstdeployed application that relaxes this assumption infavor of a bounded rationality model called SUQR,which models the adversary’s stochastic response todefender’s strategy (Nguyen et al. 2013). SUQR wasshown to perform the best in human subject experi-ments when compared with other models. SUQR pre-dicts the adversary’s probability of attacking i basedon a linear combination of three key features at thetargets, including the coverage probability ci, theattacker’s reward Uar,i and penalty U

ap,i. A set of param-

eters (w1, w2, w3) are used for combining the features.They indicate the importance of the features and canbe learned from data.

First Tests and FeedbackWe first tested PAWS-Initial (Yang et al. 2014) atUganda’s Queen Elizabeth National Park (QENP) forthree days. Subsequently, with the collaboration ofPanthera and Rimba, we started working in forests inMalaysia in September 20141. These protected forestsare home to endangered animals such as the Malayantiger and Asian elephant but are threatened bypoachers. One key difference of this site compared to

QENP is that there are large changes in elevation, andthe terrain is much more complex. The first four-daypatrol in Malaysia was conducted in November 2014.For this test, we set the value of Uar,i (input for PAWS-Initial) by aggregating observation data recorded dur-ing April 2014–September 2014. We first set theimportance value of each cell as a weighted sum ofthe observed counts of different types of animals anddifferent human activities. We then dilute the impor-tance value of available cells to nearby cells by apply-ing a 5 by 5 Gaussian kernel for a convolution oper-ation so as to estimate the value for cells with no datarecorded.

These initial tests revealed four areas of shortcom-ings, which restricted PAWS-Initial from being usedregularly and widely. The first limitation, which wassurprising given that it has received no attention inprevious work on security games, is the criticalimportance of topographic information that wasignored in PAWS-Initial. Topography can affectpatrollers’ speed in key ways. For example, lakes areinaccessible for foot patrols. Not considering suchinformation may lead to failure to complete thepatrol route. Figure 2 shows one patrol route duringthe test in Uganda. The suggested route (orangestraight line) goes across the water body (lower rightpart of figure), and hence the patrollers decided towalk along the water body (black line). Also, changesin elevation require extra patrol effort, and extremechanges may stop the patrollers from following aroute. For example, in figure 3a, PAWS-Initialplanned a route on a 1 kilometer by 1 kilometer grid(straight lines), and suggested that the patrollers walkto the north area (row 1, column 3) from the southside (row 2, column 3). However, such movementwas extremely difficult because of the changes in ele-vation. So patrollers decided to head toward the

Figure 1. Snares Found by Patrollers.

northwest area as the elevation change is more gen-tle. In addition, it is necessary to focus on terrain fea-tures such as ridgelines and streams (figure 3b) whenplanning routes for three reasons:

First, they are important conduits for certain mam-mal species such as tigers; hence, second, poachersuse these features for trapping and moving about ingeneral; and third, patrollers find it easier to movearound here than on slopes. Figure 4a shows a promi-nent ridgeline.

The second limitation is that PAWS-Initial assumesthe payoff values of the targets — for example, Uar,i —are known and fixed. In the domain of wildlife pro-tection, there can be uncertainties due to animalmovement and seasonal changes. Thus, consideringpayoff uncertainty is necessary for optimizing patrolstrategy.

The third limitation is that PAWS-Initial cannotscale to provide detailed patrol routes in large con-servation areas, which is necessary for successfuldeployment. Detailed routes require fine-grained dis-cretization, which leads to an exponential number ofroutes in total.

The fourth limitation is that PAWS-Initial consid-ers covering individual grid cells, but not feasibleroutes. In practice, the total patrolling time is limit-ed, and the patrollers can move to nearby areas. Apatrol strategy for implementation should be in theform of a distribution over feasible patrol routes sat-isfying these constraints. Without taking thesescheduling (routing) constraints into account, theoptimal coverage probabilities calculated by PAWS-Initial may not be implementable. Figure 4b showsan example area that is discretized into four cells andthe base camp is located in the upper left cell. There

are three available patrol routes, each protecting twotargets. The coverage probabilities shown in figure 4ccannot be achieved by a randomization over thethree routes because the coverage of the upper leftcell (Target 1) should be no less than the overall cov-erage of the remaining three cells since all routes startfrom the base camp.

PAWS OverviewFigure 5 provides an overview of the deployed ver-sion of PAWS. PAWS first takes the input data andestimates the animal distribution and human activi-ty distribution. Based on this information, an SSG-based game model is built, and the patrol strategy iscalculated. In wildlife protection, there is repeatedinteraction between patrollers and poachers. Whenpatrollers execute the patrol strategy generated byPAWS over a period (for example, three months),more information is collected and can become part ofthe input in the next round.

PAWS provides significant innovations in address-ing the aforementioned limitations of PAWS-Initial.In building the game model, PAWS uses a novel hier-archical modeling approach to building a virtualstreet map, while incorporating detailed topographicinformation. PAWS models the poachers boundedrationality as described by the SUQR model and con-siders uncertainty in payoff values. In calculating thepatrol strategy, PAWS uses the ARROW (Nguyen et al.2015) algorithm to deal with payoff uncertainty andadopts cutting plane approach and column genera-tion to address the scalability issue introduced byscheduling constraints.

Input and Initial AnalysisThe input information includes contour lines thatdescribe the elevation, terrain information such aslakes and drainage, base camp locations, previousobservations (animals and human activities), as wellas previous patrol tracks. However, the point detec-tions of the presence of animal and human activityare not likely to be spatially representative. As such,it is necessary to predict the animal and human activ-ity distribution over the entire study area. To thisend, we used (1) JAGS (Plummer 2003) to produce aposterior predictive density raster for tigers (as a tar-get species) derived from a spatially explicit capture-recapture analysis conducted in a Bayesian frame-work; and (2) MaxEnt (Phillips, Anderson, andSchapire 2006) to create a raster of predicted humanactivity distribution based on meaningful geograph-ical covariates (for example, distance to water, slope,elevation) in a maximum entropy modeling frame-work.

Build Game ModelBased on the input information and the estimateddistribution, we build a game model abstracting the

Articles

26 AI MAGAZINE

Figure 2. One Patrol Route During the Test in Uganda.

strategic interaction between the patroller and thepoacher as an SSG. Building a game model involvesdefender action modeling, adversary action model-ing, and payoff modeling. We will discuss all threeparts but emphasize defender action modeling sincethis is one of the major challenges to bring PAWS toa regularly deployed application. Given the topo-graphic information, modeling defender actions inPAWS is far more complex than any other previoussecurity game domain.

Defender Action Modeling Based on the feedback from the first tests, we aim toprovide detailed guidance to the patrollers. If we usea fine-grained grid and treat every fine-grained gridcell as a target, computing the optimal patrollingstrategy is exceptionally computationally challeng-ing due to the large number of targets and the expo-nential number of patrol routes. Therefore, a keynovelty of PAWS is to provide a hierarchical model-ing solution, the first such model in security gameresearch. This hierarchical modeling approach allowsus to attain a good compromise between scaling upand providing detailed guidance. This approachwould be applicable in many other domains for largeopen area patrolling where security games are appli-cable, not only other green security games applica-tions, but others including patrolling of large ware-house areas or large open campuses by robots orUAVs.

More specifically, we leverage insights from hierar-chical abstraction for heuristic search such as pathplanning (Botea, Müller, and Schaeffer 2004) andapply two levels of discretization to the conservationarea. We first discretize the conservation area into 1kilometer by 1 kilometer grid cells and treat everygrid cell as a target. We further discretize the grid cellsinto 50 meters by 50 meters raster pieces and describethe topographic information such as elevation in 50-meter scale. The defender actions are patrol routesdefined over a virtual “street map” — which is builtin terms of raster pieces while aided by the grid cellsin this abstraction as described below. With this hier-archical modeling, the model keeps a small numberof targets and reduces the number of patrol routeswhile allowing for details at the 50-meter scale.

The street map is a graph consisting of nodes andedges, where the set of nodes is a small subset of theraster pieces, and edges are sequences of raster pieceslinking the nodes. We denote nodes as key accesspoints (KAPs) and edges as route segments. The streetmap not only helps scalability but also allows us tofocus patrolling on preferred terrain features such asridgelines. The street map is built in three steps: (1)determine the accessibility type for each raster piece,(2) define KAPs, and (3) find route segments to linkthe KAPs.

In the first step, we check the accessibility type ofevery raster piece. For example, raster pieces in a lakeare inaccessible, whereas raster pieces on ridgelines or

previous patrol tracks are easily accessible. Ridgelinesand valley lines are inferred from the contour linesusing existing approaches in hydrology (Tarboton,Bras, and Rodriguez-Iturbe 2007).

The second step is to define a set of KAPs, throughwhich patrols will be routed. We want to build thestreet map in such a way that each grid cell can bereached. So we first choose raster pieces that canserve as entries and exits for the grid cells as KAPs,that is, the ones that are on the boundary of grid cellsand are easily accessible according to the accessibili-ty type calculated in the first step. When there are

Articles

SPRING 2017 27

Figure 3. First Four-Day Patrol in Malaysia.

Figure 3a shows one suggested route (orange straight lines) and the actualpatrol track (black line). Figure 3b shows the patrollers walking along thestream during the patrol.

A

B

multiple adjacent raster pieces that are all easilyaccessible, we add the midpoint as a KAP. If there areno easily accessible raster pieces on one side of theboundary, we choose the raster piece with the lowestslope as a KAP. In addition, we consider existing basecamps as KAPs as they are key points in planning thepatroller’s route. We choose additional KAPs toensure KAPs on the boundary of adjacent cells arepaired. Figure 6 shows identified KAPs and easilyaccessible pieces (black and gray raster pieces respec-tively).

The last step is to find route segments to connectthe KAPs. Instead of inefficiently finding route seg-ments to connect each pair of KAPs on the map glob-ally, we find route segments locally for each pair ofKAPs within the same grid cell, which is sufficient toconnect all the KAPs. When finding the route seg-ment, we design a distance measure that estimatesthe actual patrol effort and also gives high priority tothe preferred terrain features. The effort needed forthree-dimensional movement can be interpreted asthe equivalent distance on flat terrain. For example,for gentle slopes, equivalent “flat-terrain” distance isobtained by adding eight kilometers for every onekilometer of elevation ascent according to Naismith’srule (Thompson 2011). In PAWS, we apply Naismith’srule with Langmuir corrections (Langmuir 1995) forgentle slopes (< 20°) and apply Tobler’s hiking speedfunction (Tobler 1993) for steep slopes (≥ 20°). Verysteep slopes (> 30°) are not allowed. We penalize notwalking on preferred terrain features by adding extradistance. Given the distance measure, the route seg-ment is defined as the shortest distance path linkingtwo KAPs within the grid cell.

The defender’s pure strategy is defined as a patrolroute on the street map, starting from the base camp,walking along route segments and ending with basecamp, with its total distance satisfying the patrol dis-tance limit (all measured as the distance on flat ter-rain). The patroller confiscates the snares along theroute and thus protects the grid cells. More specifi-cally, if the patroller walks along a route segment thatcovers a sufficiently large portion (for example, 50percent of animal distribution) of a grid cell, the cellis considered to be protected. The defender’s goal is tofind an optimal mixed patrol strategy — a probabili-ty distribution over patrol routes.

Poacher Action Modeling and Payoff ModelingThe poacher’s actions are defined over the grid cellsto aid scalability. In this game, we assume the poach-er can observe the defender’s mixed strategy and thenchooses one target (a grid cell) and places snares inthis target. Following earlier work, the poacher in thisgame is assumed to be boundedly rational, and hisactions can be described by the SUQR model.

Each target is associated with payoff values indi-cating the reward and penalty for the patrollers andthe poachers. As mentioned earlier, PAWS models azero-sum game and the reward for the attacker (and

Articles

28 AI MAGAZINE

Figure 4. Illustrative Examples.

a. Ridgeline. b. Feasible routes. c. Coverage

A

Route 3

Route 3 Route 2

B

0.8 0.2

0.5 0.5

C

the penalty for the defender) is decided by the animaldistribution. However, in this game model, we needto handle uncertainty in the players’ payoff valuessince key domain features, such as animal density,that contribute to the payoffs are difficult to precise-ly estimate. In addition, seasonal or dynamic animalmigration may lead to payoffs to become uncertainin the next season. We use intervals to represent pay-off uncertainty in PAWS; the payoffs are known to liewithin a certain interval whereas the exact values areunknown. Interval uncertainty is, in fact, a well-known concept to capture uncertainty in securitygames (Nguyen et al. 2014, 2015). We determine thesize of the payoff intervals at each grid cell based onpatrollers’ patrol efforts at that cell. If the patrollerspatrol a cell more frequently, there is less uncertain-ty in the players’ payoffs at that target and thus asmaller size of the payoff intervals.

Calculate Patrol StrategyWe build on algorithms from the rich security gameliterature to optimize the defender strategy. However,we find that no existing algorithm directly fits our

needs as we need an algorithm that can scale up tothe size of the domain of interest, where (1) we mustgenerate patrol routes over the street map over theentire conservation area region, while (2) simultane-ously addressing payoff uncertainty and (3) bound-ed rationality of the adversary. While the ARROW(Nguyen et al. 2015) algorithm allows us to address(2) and (3) together, it cannot handle scale-up overthe street map. Indeed, while the (virtual) street mapis of tremendous value in scaling up as discussed ear-lier, scaling up given all possible routes (approxi-mately equal to 1012 routes) on the street map is stilla massive research challenge. We, therefore, inte-grate ARROW with another algorithm BLADE (Yanget al. 2013) for addressing the scalability issue, result-ing in a novel algorithm that can handle all the threeaforementioned challenges. The new algorithm isoutlined in figure 7. In the following, we explainhow ARROW and BLADE are adapted and integrated.

ARROW attempts to compute a strategy that isrobust to payoff uncertainty given that poachers’responses follow SUQR. The concept of minimizingmaximum regret is a well-known concept in AI for

Articles

SPRING 2017 29

Figure 5. PAWS Overview.

Calculate PatrolStrategy

Build GameModel

Initial Analysis

Input Terrain

Information

Patrol Track Observation Data

Street Map Payoff UncertaintyPoacher Behavior

Model

ARROW

Human ActivityDistribution

AnimalDistribution

Route Generation

Cutting Plane Exec

ute

and

Col

lect

Obs

erva

tions

decision making under uncertainty (Wang andBoutilier 2003). ARROW uses the solution concept ofbehavioral minimax regret to provide the strategythat minimizes regret or utility loss for the patrollersin the presence of payoff uncertainty and boundedrational attackers. In small-scale domains, ARROWcould be provided all the routes (the defender’s purestrategies), on the basis of which it would calculatethe PAWS solution — a distribution over the routes.Unfortunately, in large-scale domains like ours, enu-merating all the routes is infeasible. We must, there-fore, turn to an approach of incremental solutiongeneration, which is where it interfaces with theBLADE framework.More specifically, for scalability reasons, ARROW firstgenerates the robust strategy for the patrollers in theform of coverage probabilities over the grid cellswithout consideration of any routes. Similar toBLADE, a separation oracle is then called to check ifthe coverage vector is implementable. If it is imple-mentable, the oracle returns a probability distribu-tion over patrol routes that implements the coveragevector, which is the desired PAWS solution. If it is notimplementable (see figure 4c for an example of a cov-erage vector that is not implementable) the oraclereturns a constraint (cutting plane) that informsARROW why it is not. For the example in figure 4band 4c, if ARROW generates a vector as shown in fig-ure 4c, the constraint returned could be

since all implementable coverage vectors should sat-isfy this constraint. This constraint helps ARROWrefine its solution. The process repeats until the cov-erage vector generated by ARROW is implementable.

As described in BLADE (Yang et al. 2013), to avoidenumerating all the feasible routes to check whetherthe coverage vector is implementable, the separationoracle iteratively generates routes until it has justenough routes (usually after a small number of itera-tions) to match the coverage vector probabilities orget the constraint (cutting plane). At each iterationof route generation (shown in the bottommost boxin figure 7), the new route is optimized to cover tar-gets of high value. However, we cannot directly useany existing algorithm to find the optimal route ateach iteration due to the presence of our street map.But we note similarities to the well-studied orienteer-ing problem (Vansteenwegen, Souffriau, and Oud-heusden 2011) and exploit the insight of the S-algo-rithm for orienteering (Tsiligiridis 1984).

In particular, in this bottommost box in figure 7, toensure each route returned is of high quality, we runa local search over a large number of routes andreturn the one with the highest total value. In everyiteration, we start from the base KAP and choosewhich KAP to visit next through a weighted randomselection. The next KAP to be visited can be any KAPon the map, and we assume the patroller will take theshortest path from the current KAP to the next KAP.The weight of each candidate KAP is proportional tothe ratio of the additional target value that can beaccrued and distance from the current KAP. We setthe lower bound of the weight to be a small positivevalue to make sure every feasible route can be chosenwith positive probability. The process continues untilthe patroller has to go back to the base to meet thepatrol distance limit constraint. Given a large num-ber of such routes, our algorithm returns a route closeto the optimal solution.

Integrating all these algorithms, PAWS calculatesthe patrol strategy consisting of a set of patrol routesand the corresponding probability for taking them.

Addressing Additional Practical Challenges

We have introduced the technical innovations thatlead to PAWS’s deployment. In addition to theseinnovations, we have addressed a number of practi-cal constraints to make the strategy suggested byPAWS easier to follow by human patrollers. In thissection, we summarize these challenges and our solu-tions to them.

First, mountaintops should be considered as keypoints in the patrol route. In PAWS, we require thatthe patrollers always move between KAPs, which arelocated at the boundary of the grid cells or the base

c1 ! cii=24

"

Articles

30 AI MAGAZINE

Figure 6. KAPs (Black) for 2 by 2 Grid Cells.

camps. Therefore, in some suggested patrol routes,the patroller is asked to go to a mountaintop and godownhill for a short distance and then backtrack.However, this kind of short downhill followed byreturning uphill will annoy the patrollers, and theywill naturally ignore the downhill part. We addressthis problem by considering mountaintops as KAPswhen building the street map. With these additionalKAPs, patrollers are not forced to take the shortdownhill unless necessary (that is, when the shortdownhill covers areas with high animal density).

Second, there is a limit on working time in addi-tion to the limit on walking distance. It takes lesstime for the patroller to go to an area and then back-track than to take a loop even if the walking distanceis the same. The reason is the patrollers need to spendthe time to record what they observe, including ani-mal signs and human activity signs. If the patrollerswalk along the same ridgeline twice in a day, theyonly need to record the signs once. Therefore, indesigning the patrol routes, we should consider thetotal working time in addition to the total distance.This is implemented in PAWS by adding additionalconstraints in route generation.

Third, not all terrain features should be treated inthe same way. In building the street map, we givepreference to terrain features like ridgelines bydesigning a distance measure that penalizes not walk-ing along the terrain feature. However, how muchpriority should be given to the terrain featuresdepends on the cost of the alternative routes, or how

much easier it is compared to taking other routes. Onthe one hand, in a very hilly region of the area wherethere are large elevation changes, the patrollerswould highly prefer the terrain features as it is mucheasier to walk along them than taking an alternativeroute. On the other hand, if the elevation change inthe region is small, the effort of taking a ridgeline forunit distance is comparable to that of taking an alter-native route. To differentiate these different cases, weuse secondary derivatives to check how importantthe ridgeline is. Instead of penalizing not walkingalong the terrain features, we can use a discount fac-tor for taking the preferred route, and assign a high-er discount factor for terrain features with a higher(regarding absolute value) secondary derivative.

Finally, additional factors such as slope should beconsidered when evaluating the walking effort. Inthe distance measure introduced in the previous sec-tion, elevation change and terrain features have beenconsidered, but there are other factors that con-tribute to the walking effort. For example, walkingalong the contour line will lead to zero elevationchange along the way, but the effort needed highlydepends on the slope. Walking along the hillside of asteep slope takes much more effort than walking onthe flat terrain. Therefore, in the distance measure,we penalize walking along the hillside and assign ahigher penalty factor for a higher slope.

Trading Off Exploration and ExploitationBuilding on the PAWS framework, we provide a vari-

Articles

SPRING 2017 31

Figure 7. New Integrated Algorithm.

Separation Oracle

S = S ∪ s

ARROW: compute optimal coveragevector c given a set of linear constraints S.ˆ

Route Generation: !nd routes thatconstitute the separation hyperplane.

FindCuttingPlane: Find a hyperplaneseparating c and feasible region C. If exists,

c ∉ C and a new constraint s is found.ˆˆ

Is c ∈ C? ˆ

ation of PAWS (denoted as PAWS-EvE) that offers theoption of assigning a probability range for selectingan explorative route. Explorative routes are thosethat cover a significant portion of previously unpa-trolled land, while exploitative routes are those thatcover a significant portion of land previouslypatrolled.

The major advantage of selecting exploitativeroutes is that patrollers are familiar with those patrolroutes. Having experience with the routes, patrollersalso require less effort on following the routes as theywould with unfamiliar territory, enabling them tobetter cover the area they patrol. Further, the explo-rative routes may require more effort on checking themap, finding water refill points, and figuring out thebest way around unexpected obstacles. Some of thesetasks require experienced patrollers and additionalequipment, which are not available for every patrol.However, if patrollers were to only take exploitativeroutes, then poachers would easily be able to observesuch a strategy and then focus on targeting otherareas. With the objective of PAWS to minimizepoaching activity, it is necessary to also take explo-rative routes. Therefore, offering the option of settinga probability range for selecting an explorative routewould be helpful for practical use. The range is setbefore generating the patrols based on these practicalconcerns.

To implement the functionality of PAWS-EvE, wemodify the previously mentioned separation oracleby introducing two sets of routes, the explorative setand the exploitative set. Two new user-definedparameters θ and δ are introduced and we add twonew constraints to make sure the probability ofassigning a patrol route from the explorative set lieswithin a δ margin of θ. Also, in route generation, we

check if adding a route from the explorative setwould improve the solution, and then we also checkif adding a route from the exploitative set wouldimprove the solution.

Deployment and EvaluationPAWS patrols are now regularly deployed at a conser-vation area in Malaysia. This section provides detailsabout the deployment and both subjective and objec-tive evaluations of PAWS patrols.

PAWS patrol aims to conduct daily patrols frombase camps. Before the patrol starts, PAWS generatesthe patrol strategy starting from the base camp select-ed by the patrol team leader. The patrol distance lim-it considered by PAWS is 10 kilometer per day (equiv-alent flat terrain). As shown in table 1, this leads toabout 9000 raster pieces to be considered. Thus, it isimpossible to consider each raster piece as a separatetarget or consider all possible routes over the rasterpieces. With the two-level discretization and thestreet map, the problem scale is reduced, with 8.57(=194.33/22.67) KAPs and 80 route segments in eachgrid cell on average, making the problem manage-able. The strategy generated by PAWS is a set of sug-gested routes associated with probabilities and theaverage number of suggested routes associated withprobability > 0.001 is 12.

Each PAWS patrol lasts for 4–5 days and is execut-ed by a team of 3–7 patrollers. The patrol planner willmake plans based on the strategy generated by PAWS.After reaching the base camp, patrollers execute dai-ly patrols, guided by PAWS’s patrol routes. Table 2provides a summary of basic statistics about thepatrols. During the patrol, the patrollers are equippedwith a printed map, a hand-held GPS, and data-recording booklet. They detect animal and humanactivity signs and record them with detailed com-ments and photos. After the patrol, the data manag-er will put all the information into a database, includ-ing patrol tracks recorded by the hand-held GPS, andthe observations recorded in the logbook.

Figure 8 shows various types of signs found duringthe patrols. Table 3 summarizes all the observations.These observations show that there is a serious ongo-ing threat from the poachers. Column 2 shows resultsfor all PAWS patrols. Column 3 shows results forexplorative PAWS patrols, the (partial) patrol routes,which go across areas where the patrollers have nev-er been before. To better understand the numbers, weshow in column 4 the statistics about early-stagenon-PAWS patrols in this conservation area, whichwere deployed for a tiger survey. Although it is not afair comparison as the objectives of the non-PAWSpatrols and PAWS patrols are different, comparingcolumns 2 and 3 with column 4 indicates that PAWSpatrols are effective in finding human activity signsand animal signs. Finding the human activity signs isimportant to identify hotspots of poaching activity,

Articles

32 AI MAGAZINE

Table 1. Problem Scale for PAWS Patrols.

Average # of Reachable Raster Pieces 9066.67

Average # of Reachable Grid Cells (Targets) 22.67

Average # of Reachable KAPs 194.33

Table 2. Basic Information of PAWS Patrols.

Average Trip Length 4.67 Days

Average Number of Patrollers 5

Average Patrol Time Per Day 4.48 hours

Average Patrol Distance Per Day 9.29 km

and patrollers’ presence will deter the poachers. Ani-mals signs are not a direct evaluation of PAWSpatrols, but they indicate that PAWS patrols prioritizeareas with higher animal density. Finding these signsis aligned with the goal of PAWS — combat poachingto save animals — and thus is a proof for the effec-

tiveness of PAWS. Comparing column 3 with column2, we find the average number of observations madealong the explorative routes is comparable to andeven higher than that of all PAWS patrol routes. Theobservations on explorative routes are important asthey lead to a better understanding of the unex-

Articles

SPRING 2017 33

Figure 8. Various Signs Recorded During PAWS Patrols.

A B

C D

Table 3. Summary of Observations.

Patrol Type All PAWS Patrol Explorative PAWS Patrol Previous Patrol for Tiger Survey

Total Distance (kilometers) 130.11 20.1 624.75 Average Number of Human Activity Signs per kilometer

0.86 1.09 0.57

Average Number of Animal Signs per kilometer

0.41 0.44 0.18

Courtesy Uda Alok

plored area. These results show that PAWS can guidethe patrollers toward hotspots of poaching activityand provide valuable suggestions to the patrol plan-ners.

Along the way of PAWS deployment, we havereceived feedback from patrol planners andpatrollers. The patrol planners mentioned that thetop routes in PAWS solution (routes with the highestprobability) come close to an actual planner’s routes,which shows PAWS can suggest feasible routes andpotentially reduce the burden of the planning effort.As we deploy PAWS in the future at other sites, thecumulative human planners’ effort saved by usingPAWS will be a considerable amount. In addition,patrollers commented that PAWS was able to guidethem toward poaching hotspots. The fact that theyfound multiple human signs along the explorativePAWS patrol routes makes them believe that PAWS isgood at finding good ridgelines that are taken by ani-mals and humans. Patrollers and patrol planners alsoagree that PAWS generates detailed suggested routes,which can guide the actual patrol. Patrollers com-mented that the suggested routes were mostly alongthe ridgeline, which is easier to follow, comparedwith the routes from the first trial by PAWS-Initial.Figure 9 shows one suggested route (orange line) andthe actual patrol track (black line) during PAWSpatrol in August 2015 (shown on a 1 kilometer by 1kilometer grid). Due to the precision of the contourlines we get, we provide a 50-meter buffer zone (lightorange polygon) around the suggested route(orange/light-gray line). The patrollers started fromthe base camp (the green or shaded triangle) andheaded to the southeast. The patrollers mostly fol-lowed PAWS’s suggested route, indicating that theroute generated by PAWS is easy to follow (contrast

with PAWS-Initial as shown in figure 3a). Finally, thepower of randomization in the PAWS solution can beexpected in the long term.

Lessons LearnedDuring the development and deployment process,we faced several challenges, and here we outlinesome lessons learned.

First, firsthand immersion in the security environ-ment of concern is critical to understanding the con-text and accelerating the development process. Theauthors (from USC and NTU) intentionally went forpatrols in the forest with the local patrolling team tofamiliarize themselves with the area. The firsthandexperience confirmed the importance of ridgelines,as several human and animal signs were found alongthe way, and also confirmed that extreme changes inelevation require a considerable extra effort of thepatrollers. This gave us the insight for building thestreet map.

Second, visualizing the solution is important forcommunication and technology adaptation. Whenwe communicate with domain experts and humanplanners, we need to effectively convey the game-theoretic strategy generated by PAWS, which is aprobability distribution over routes. We first visualizethe routes with probability > 0.01 using ArcGIS sothat they can be shown on the topographic map andthe animal distribution map. Then for each route, weprovide detailed information that can assist thehuman planners’ decision making. We not only pro-vide basic statistics such as probability to be takenand total distance, but also estimate the difficulty lev-el for patrol, predict the probability of finding ani-mals and human signs, and provide an elevationchart that shows how the elevation changes alongthe route. Such information can help planners’understanding of the strategy, and also help the plan-ner assign patrol routes to the appropriate team ofpatrollers, as some patrollers may be good at long-dis-tance walking with flat terrain while others wouldprefer short-distance hiking with high elevationchange.

Third, minimizing the need for extra equip-ment/effort would further ease PAWS future deploy-ment, that is, patrollers would prefer having a singlehand-held device for collecting patrol data and dis-playing suggested patrol routes. If PAWS routes couldbe embedded in the software that is already in use forcollecting data in many conservation areas, for exam-ple, SMART, it would reduce the effort required ofplanners. This is one direction for future develop-ment.

SummaryPAWS is a first deployed green security game applica-tion to optimize human patrol resources to combat

Articles

34 AI MAGAZINE

Figure 9. One Daily PAWS Patrol Route in August 2015.

poaching. We provided key research advances toenable this deployment; this has provided a practicalbenefit to patrol planners and patrollers. The deploy-ment of PAWS patrols will continue at the site inMalaysia. Panthera has seen the utility of PAWS, andwe are taking steps to expand PAWS to its other sites.This future expansion and maintenance of PAWS willbe taken over by Armorway,3 a security games com-pany (starting in spring 2016); Armorway has signif-icant experience in supporting security-games-basedsoftware deployments.

Acknowledgement

This research was supported by MURI GrantW911NF-11-1-0332. We also thank our partners inthe field who made these tests possible and subcon-tract from Cornell University for NSF Grant CCF-1522054.

Notes1. For the security of animals and patrollers, no latitude/lon-gitude information is presented in this article.

2. The Spatial Monitoring and Reporting Tool (SMART),www.smartconservationsoftware.org.

3. The company has now changed its name. Seeavataai.com.

ReferencesBotea, A.; Müller, M.; and Schaeffer, J. 2004. Near OptimalHierarchical Path-Finding. Journal of Game Development 1(1):7–28.

Chapron, G.; Miquelle, D. G.; Lambert, A.; Goodrich, J. M.;Legendre, S.; and Clobert, J. 2008. The Impact on Tigers ofPoaching Versus Prey Depletion. Journal of Applied Ecology45(6): 1667–1674.

Fang, F.; Jiang, A. X.; and Tambe, M. 2013. Optimal PatrolStrategy for Protecting Moving Targets with Multiple MobileResources. In Proceedings of the 2013 International Conferenceon Autonomous Agents and Multi-Agent Systems, 957–964.Richland, SC: International Foundation for AutonomousAgents and Multiagent Systems.

Fang, F.; Stone, P.; and Tambe, M. 2015. When SecurityGames Go Green: Designing Defender Strategies to PreventPoaching and Illegal Fishing. In Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence(IJCAI). Palo Alto, CA: AAAI Press.

Hamisi, M. 2008. Identification and Mapping Risk Areas forZebra Poaching: A Case of Tarangire National Park, Tanza-nia. Ph.D. Dissertation, Thesis, ITC Faculty of Geo-Informa-tion Science and Earth Observation of the University ofTwente, Enschede, The Netherlands.

Haskell, W.; Kar, D.; Fang, F.; Tambe, M.; Cheung, S.; andDenicola, E. 2014. Robust Protection of Fisheries with Com-pass. In Proceedings of the Twenty-Eighth AAAI Conference onArtificial Intelligence and the Twenty-Sixth Innovative Applica-tions of Artificial Intelligence Conference, 2978–2983. PaloAlto, CA: AAAI Press.

IUCN. 2015. IUCN Red List of Threatened Species. Version2015.2. Gland, Switzerland: International Union for Con-servation of Nature. (www.iucnredlist.org)

Johnson, M. P.; Fang, F.; and Tambe, M. 2012. Patrol Strate-

gies to Maximize Pristine Forest Area. In Proceedings of theTwenty-Sixth AAAI Conference on Artificial Intelligence (AAAI),295–301. Palo Alto, CA: AAAI Press.

Kar, D.; Fang, F.; Fave, F. D.; Sintov, N.; and Tambe, M. 2015.A Game of Thrones: When Human Behavior Models Com-pete in Repeated Stackelberg Security Games. In Proceedingsof the International Conference on Autonomous Agents andMultiagent Systems. Richland, SC: International Foundationfor Autonomous Agents and Multiagent Systems.

Korzhyk, D.; Conitzer, V.; and Parr, R. 2010. Complexity ofComputing Optimal Stackelberg Strategies in SecurityResource Allocation Games. In Proceedings of the 24thNational Conference on Artificial Intelligence, 805–810. PaloAlto, CA: AAAI Press.

Langmuir, E. 1995. Mountaincraft and Leadership: A Hand-book for Mountaineers and Hillwalking Leaders in the BritishIsles. Manchester, UK: Mountain Leader Training Board.

Lemieux, A. M., ed. 2014. Situational Prevention of Poaching.Crime Science Series. New York: Taylor & Francis Group /Routledge.

Nguyen, T. H.; Fave, F. M. D.; Kar, D.; Lakshminarayanan, A.S.; Yadav, A.; Tambe, M.; Agmon, N.; Plumptre, A. J.; Dri-ciru, M.; Wanyama, F.; and Rwetsiba, A. 2015. Making theMost of Our Regrets: Regret-Based Solutions to Handle Pay-off Uncertainty and Elicitation in Green Security Games. InDecision and Game Theory for Security: 6th International Con-ference. Berlin: Springer.

Nguyen, T. H.; Yadav, A.; An, B.; Tambe, M.; and Boutilier,C. 2014. Regret-Based Optimization and Preference Elicita-tion for Stackelberg Security Games with Uncertainty. InProceedings of the Twenty-Eighth AAAI Conference on ArtificialIntelligence. Palo Alto, CA: AAAI Press.

Nguyen, T. H.; Yang, R.; Azaria, A.; Kraus, S.; and Tambe, M.2013. Analyzing the Effectiveness of Adversary Modeling inSecurity Games. In Proceedings of the Twenty-Seventh AAAIConference on Artificial Intelligence. Palo Alto, CA: AAAI Press.

Phillips, S. J.; Anderson, R. P.; and Schapire, R. E. 2006. Max-imum Entropy Modeling of Species Geographic Distribu-tions. Ecological Modelling 190(3–4): 231–259.

Pita, J.; Jain, M.; Marecki, J.; Ordóñez, F.; Portway, C.;Tambe, M.; Western, C.; Paruchuri, P.; and Kraus, S. 2008.Deployed ARMOR Protection: The Application of a GameTheoretic Model for Security at the Los Angeles Interna-tional Airport. In Proceedings of the 7th International JointConference on Autonomous Agents and Multiagent Systems:Industrial Track, AAMAS 2008, 125–132. Richland, SC:International Foundation for Autonomous Agents andMultiagent Systems.

Plummer, M. 2003. JAGS: A Program for Analysis ofBayesian Graphical Models Using Gibbs Sampling. Paperpresented at the 3rd International Workshop on Distrib-uted Statistical COmputing, March 20–22, Vienna, Austria.

Qian, Y.; Haskell, W. B.; Jiang, A. X.; and Tambe, M. 2014.Online Planning for Optimal Protector Strategies inResource Conservation Games. In Proceedings of the Interna-tional Conference on Autonomous Agents and Multiagent Sys-tems (AAMAS’14). Richland, SC: International Foundationfor Autonomous Agents and Multiagent Systems.

Sanderson, E.; Forrest, J.; Loucks, C.; Ginsberg, J.; Diner-stein, E.; Seidensticker, J.; Leimgruber, P.; Songer, M.; Heyd-lauff, A.; OBrien, T.; Bryja, G.; Klenzendorf, S.; and Wikra-manayake, E. 2006. Setting Priorities for the Conservationand Recovery of Wild Tigers: 2005–2015. In Tigers of the

Articles

SPRING 2017 35

World: The Science, Politics, and Conservation of Panthera tigris,chapter 9, 143–161, ed. R. Tilson and P. J. Nyhus. Amster-dam, The Netherlands: Elsevier.

Shieh, E.; An, B.; Yang, R.; Tambe, M.; Baldwin, C.; DiRen-zo, J.; Maule, B.; and Meyer, G. 2012. PROTECT: A DeployedGame Theoretic System to Protect the Ports of the UnitedStates. In Proceedings of the 11th International Conference onAutonomous Agents and Multiagent Systems, Volume 1,AAMAS’12, 13–20. Richland, SC: International Foundationfor Autonomous Agents and Multiagent Systems.

Stokes, E. J. 2010. Improving Effectiveness of ProtectionEfforts in Tiger Source Sites: Developing a Framework forLaw Enforcement Monitoring Using Mist. Integrative Zoolo-gy 5(4): 363–377.

Tambe, M. 2011. Security and Game Theory: Algorithms,Deployed Systems, Lessons Learned. Cambridge, UK: Cam-bridge University Press.

Tarboton, D. G.; Bras, R. L.; and Rodriguez-Iturbe, I. 2007.On the Extraction of Channel Networks from Digital Eleva-tion Data. Hydrologic Processes 5(1): 81–100.

Thompson, S. 2011. Unjustifiable Risk?: The Story of BritishClimbing. Cumbria, UK: Cicerone Press.

Tobler, W. 1993. Three Presentations on GeographicalAnalysis and Modeling. Nonisotropic Geographic Model-ing: Speculations on the Geometry of Geography, and Glob-al Spatial Analysis. Technical Report 93-1, National centerfor Geographic Information and Analysis. Santa Barbara,CA: University of California, Santa Barbara.

Tsiligiridis, T. 1984. Heuristic Methods Applied to Orien-teering. The Journal of the Operational Research Society 35(9):797–809.

Vansteenwegen, P.; Souffriau, W.; and Oudheusden, D. V.2011. The Orienteering Problem: A Survey. European Journalof Operational Research 209(1): 1–10.

Wang, T., and Boutilier, C. 2003. Incremental Utility Elici-tation with the Minimax Regret Decision Criterion. In Pro-ceedings of the Eighteenth International Joint Conference on Arti-ficial Intelligence, 309–318. San Francisco: MorganKaufmann Publishers.

Wato, Y. A.; Wahungu, G. M.; and Okello, M. M. 2006. Cor-relates of Wildlife Snaring Patterns in Tsavo West NationalPark, Kenya. Biological Conservation 132(4): 500–509.

Yang, R.; Ford, B.; Tambe, M.; and Lemieux, A. 2014. Adap-tive Resource Allocation for Wildlife Protection Against Ille-gal Poachers. In Proceedings of the International Conference onAutonomous Agents and Multiagent Systems (AAMAS). Rich-land, SC: International Foundation for Autonomous Agentsand Multiagent Systems.

Yang, R.; Jiang, A. X.; Tambe, M.; and Ordéñez, F. 2013. Scal-ing-Up Security Games with Boundedly Rational Adver-saries: A Cutting-Plane Approach. In Proceedings of the 23rdInternational Joint Conference on Artificial Intelligence. PaloAlto, CA: AAAI Press.

Yin, Z.; Jiang, A. X.; Johnson, M. P.; Kiekintveld, C.; Leyton-Brown, K.; Sandholm, T.; Tambe, M.; and Sullivan, J. P.2012. TRUSTS: Scheduling Randomized Patrols for FareInspection in Transit Systems. In Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence and the Twen-ty-Fourth Conference on Innovative Applications of ArtificialIntelligence, 2348–2355. Palo Alto, CA: AAAI Press.

Fei Fang is a postdoctoral fellow at the Center for Research

on Computation and Society (CRCS) at Harvard Universityand an adjunct assistant professor at Carnegie Mellon Uni-versity. Her research lies in the field of artificial intelligenceand multiagent systems, focusing on computational gametheory with applications to security and sustainabilitydomains.

Thanh H. Nguyen is a postdoctoral researcher at the Uni-versity of Michigan. Her research is in the field of artificialintelligence, multiagent systems, and machine learning,focusing on game theory with connections to optimization,behavioral modeling, and operations research.

Robert Pickles is a monitoring specialist and the head ofmonitoring in the Tiger Program in Panthera. He is also apostdoctoral research fellow at the University of Trent. Hegained his Ph.D. from the University of Kent and the Zoo-logical Society of London in 2010.

Wai Y. Lam is the technical advisor to Rimba’s Project Hari-mau Selamanya, as well as a Tiger Program scholar with Pan-thera and a master of science student at Universiti MalaysiaTerengganu. Her work involves research on spatiotemporalpatterns of people and wildlife in protected areas forantipoaching efforts, camera trap studies, and developmentof conservation technology in database management.

Gopalasamy R. Clements is an associate professor at theKenyir Research Institute in Universiti Malaysia Terengganu.He is also the cofounder of Rimba (rimbaresearch.org), anonprofit research group conducting research on threat-ened species and ecosystems in Malaysia. His project, Hari-mau Selamanya, aims to improve protection of three largecarnivores in the state of Terengganu, Peninsular Malaysia.

Bo An is an assistant professor at the School of ComputerScience and Engineering of the Nanyang Technological Uni-versity. His research interests include artificial intelligence,multiagent systems, computational game theory, and opti-mization.

Amandeep Singh is a doctoral candidate at the OID Depart-ment at Wharton School of Business. His research interestslie in the areas of digital advertising, crowdfunding, andeconometric methods.

Brian C. Schwedock is an undergraduate student at theUniversity of Southern California studying computer engi-neering and computer science. His research with TeamcoreUSC includes statistical analysis and security game models.

Milind Tambe is founding codirector of CAIS, the USC Cen-ter for AI for Society, and Helen N. and Emmett H. JonesProfessor in Engineering at the University of Southern Cal-ifornia (USC). He is a fellow of AAAI and ACM, as well asrecipient of the ACM/SIGART Autonomous Agents ResearchAward.

Andrew Lemieux is a researcher at the Netherlands Insti-tute for the Study of Crime and Law Enforcement. His mainareas of interest are the spatial and temporal distribution ofcrime, the use of technology to improve law enforcementoperations, and the utility of wildlife crime analysis for data-driven antipoaching operations in Africa and Asia.

Articles

36 AI MAGAZINE