Maritime Threat Detection using Plan Recognition

Bryan Auslander1, Kalyan Moy Gupta1, & David W. Aha2

1
Knexus Research Corp.; 163 Waterfront Street; National Harbor, MD 20745
2
Navy Center for Applied Research in Artificial Intelligence; Naval Research Laboratory, Code 5514;
4555 Overlook Ave. SW; Washington, DC 20375
[email protected] | [email protected]

Abstract. Existing algorithms for maritime threat detection recognition algorithms are better suited for explaining threats
employ a variety of normalcy models that are probabilistic than are opaque and non-intuitive statistical models.
and/or rule-based. Unfortunately, they can be limited in their Plan recognition algorithms have been used for a variety of
ability to model the subtlety and complexity of multiple vessel tasks, such as detection of anomalous situation in an assisted
types and their spatio-temporal events, yet their representation is
care facility [4] and detection of terrorist threats [5]. However,
needed to accurately detect anomalies in maritime scenarios. To
address these limitations, we apply plan recognition algorithms they have rarely been applied to maritime threat recognition,
for maritime anomaly detection. In particular, we examine and not empirically compared with conventional behavior
hierarchical task network (HTN) and case-based algorithms for recognition algorithms for threat detection.
plan recognition, which detect anomalies by generating expected We address these issues as follows. We apply two plan
behaviors for use as a basis for threat detection. We compare recognition algorithms; the first uses hierarchical task
their performance with a behavior recognition algorithm on networks (HTNs), while the other uses case-based planning.
simulated riverine maritime traffic. On a set of simulated These algorithms apply consistency-based plan recognition
maritime scenarios, these plan recognition algorithms techniques [6]. We apply them to four simulated maritime
outperformed the behavior recognition algorithm, except for one
scenarios and compare their performance versus a Markov
reactive behavior task in which the inverse occurred.
Furthermore, our case-based plan recognizer outperformed our logic network (MLN) algorithm, a probabilistic behavior
HTN algorithm. On the short-term reactive planning scenarios, recognizer that performed well in our prior studies on
the plan recognition algorithms outperformed the behavior maritime threat detection [3]. Our results show that the plan
recognition algorithm on routine plan following. However, they recognition algorithms outperform this probabilistic behavior
are significantly outperformed on the anomalous scenarios. recognizer for scenarios involving longer-term plans.
However, the results on scenarios with reactive plans
involving anomaly situations are mixed.
I. INTRODUCTION
We structure the remainder of this paper as follows. We
Early detection and neutralization of threats from small boats provide an overview of plan recognition approaches followed
is a critical requirement for the US Navy. Detection of small by a detailed description of HTN and case-based approaches.
boat threats is particularly challenging in busy ports, harbors We then describe our empirical evaluation and results. Finally,
and riverine areas because they operate in close proximity to we conclude with a discussion and issues for future research.
large but much less maneuverable vessels. Various approaches
for automated threat and anomaly detection have been II. PLAN RECOGNITION OVERVIEW
developed to address this problem. For example, perimeter Plan recognition is the task of inferring plan(s) of an
breach detection algorithms (e.g., [1][2]) trigger an alarm intelligent agent by observing the agent’s actions or the effects
when a distance threshold is crossed. However, these can lead of those actions. It involves mapping a temporal sequence of
to numerous false alarms and unacceptable operating observed actions to some plan representation that identifies
requirements in narrow traffic lanes. To address this problem, the plan’s goal and the relation of actions among the plan’s
we recently showed that some behavior-based threat detection components [7]. Plan recognition algorithms can be
algorithms, based on probabilistic graphical models, can categorized into consistency-based and probabilistic
outperform perimeter breach algorithms [3]. However, approaches [8]. The former include hypothesize and revise
perimeter and behavior-based approaches rely on learning algorithms, version space techniques, and other closed-world
normalcy models that predominately focus on low level reasoning algorithms, while probabilistic algorithms include
activities, but cannot detect anomalies that require knowledge those that use stochastic grammars and probabilistic relational
of planned routes and schedules. models. The maritime environment is a continuous, non-
In this paper, we apply plan recognition algorithms to deterministic domain, making it challenging to apply any plan
continuously monitor activities and identify the plans of recognition approach. As a first attempt, we consider only
maritime vessels in advance of any threat occurrence. We consistency-based approaches.
assume that, given a plan, we can identify its unique goal, and
conjecture that identifying goals could help to detect III. MARITIME PLAN RECOGNITION
threatening and anomalous situations earlier and more We develop consistency-based plan recognition algorithms for
accurately. Also, the alerts that are generated by plan the maritime domain as follows (see Figure 1). We assume
 Form Approved
Report Documentation Page OMB No. 0704-0188

Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and
maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington
VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it
does not display a currently valid OMB control number.

1. REPORT DATE 3. DATES COVERED

2. REPORT TYPE
NOV 2012 00-00-2012 to 00-00-2012
4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER
Maritime Threat Detection using Plan Recognition 5b. GRANT NUMBER

5c. PROGRAM ELEMENT NUMBER

6. AUTHOR(S) 5d. PROJECT NUMBER

5e. TASK NUMBER

5f. WORK UNIT NUMBER

7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION

REPORT NUMBER
Knexus Research Corp,163 Waterfront Street,National
Harbor,MD,20745
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)

11. SPONSOR/MONITOR’S REPORT

NUMBER(S)

12. DISTRIBUTION/AVAILABILITY STATEMENT

Approved for public release; distribution unlimited
13. SUPPLEMENTARY NOTES
Proceedings of the Twelfth Annual IEEE Conference on Technologies for Homeland Security, 13-15 Nov
2012, Greater Boston, MA, pp. 249-254. U.S. Government or Federal Rights License
14. ABSTRACT
Existing algorithms for maritime threat detection employ a variety of normalcy models that are
probabilistic and/or rule-based. Unfortunately, they can be limited in their ability to model the subtlety
and complexity of multiple vessel types and their spatio-temporal events, yet their representation is needed
to accurately detect anomalies in maritime scenarios. To address these limitations, we apply plan
recognition algorithms for maritime anomaly detection. In particular, we examine hierarchical task
network (HTN) and case-based algorithms for plan recognition, which detect anomalies by generating
expected behaviors for use as a basis for threat detection. We compare their performance with a behavior
recognition algorithm on simulated riverine maritime traffic. On a set of simulated maritime scenarios,
these plan recognition algorithms outperformed the behavior recognition algorithm, except for one
reactive behavior task in which the inverse occurred. Furthermore, our case-based plan recognizer
outperformed our HTN algorithm. On the short-term reactive planning scenarios the plan recognition
algorithms outperformed the behavior recognition algorithm on routine plan following. However, they are
significantly outperformed on the anomalous scenarios.
15. SUBJECT TERMS

16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF
ABSTRACT OF PAGES RESPONSIBLE PERSON
a. REPORT b. ABSTRACT c. THIS PAGE Same as 6
unclassified unclassified unclassified Report (SAR)
Standard Form 298 (Rev. 8-98)
Prescribed by ANSI Std Z39-18
that an agent controlling a small boat or ferrry can sense the In a dynamic maritime environ nment, it is necessary to
environment, maintain beliefs about the w world state, and account for plan deviations in the face of unexpected
derive a plan (i.e., a series of actions, such as turn, cruise, and situations. For example, a vessel may m temporarily change its
ram) to accomplish a given goal (e.g., destroyy a Navy vessel). course to avoid a collision with ano other oncoming vessel. To
We assume the recognizer has access to a library of plans account for such local variations, when the observed and
(e.g., create a terrorist event, catch fish, or ferrry passengers). It expected activities deviate, the Recognizer
R replans in an
observes the world state and the actions of thee agent to predict attempt to explain the deviation. First, it updates the active
the next or some future activity. The recognizzer compares the plan based on the current state and evaluates
e it for consistency.
observations resulting from those actiions with the If the consistency check fails, it reaattempts to find a consistent
observations expected from executing its preddicted activity. If plan using a preset number of histo orical states. If the updated
these observations are consistent with one or m more of the plans plan remains inconsistent, it is i removed from future
from the library, it repeats its inference proceess. Otherwise, it consistency evaluations. Thereforee, at the start of a plan
flags the vessel as a potential threat or anom maly and presents recognition process, the Recognizzer experiences its largest
the information to a decision-maker along withh an explanation. computational load as all the plaans must be evaluated for
In the following subsections we describe pplan recognition consistency. Its search load reducees as inconsistent plans are
algorithms that use hierarchical task networkss (HTN) or case- discarded. By default, if no plan remains
r consistent then the
based reasoning (CBR) methods to implementt these steps. situation is deemed anomalous.
B. Case-Based Planning
Case-based reasoning (CBR) is a prrocess for solving problems
by retrieving and reusing solution ns from problems that are
similar to the problem at hand [10].. A CBR system relies on a
memory of problem-solution pairs called cases. These cases
can be automatically learned fro om observations or hand-
engineered, depending on the appllication. The CBR process
includes an algorithm for case retrrieval, which involves case
similarity assessment.
We adopt CBR for consistency--based plan recognition as
Figure 1. Consistency based approach to maritimee plan recognition
follows. We manually engineered two case bases that cover
A. HTN Plan Recognition selected scenarios of interest. The first case base, called the
HTN planners operate by recursively decompposing high-level Primary Case Base (PCB), covers the entire travel path of an
tasks into lower level components. Thiss decomposition agent. The second case base, calledd the Secondary Case Base
process results in a series of low-level actions or primitive (SCB), only covers the local variattions (e.g., path deviations)
tasks, constituting a plan [9]. We adapted an HTN planner for to account for collision avoidancee among two vessels. The
consistency-based plan recognition. reason for using two case bases is that, unlike the HTN
We begin by providing the HTN Recognizer with access to algorithm, which uses a single rep presentation for plans, the
all the routine goals or intents that an agent ccould follow. For CBR algorithm cannot effectively account
a for unforeseen local
each intention, the Recognizer generates a pllan comprising a deviations from the plan using a sin
ngle case representation.
set of activities that the agent would perform if it tried to TABLE 2. PRIMARY CASE REPRESENTATION
R
pursue it. Each plan is a travel path compprising a list of Feature Description
waypoints (i.e., locations the vessel wouldd travel through) Vessel-Type Type of veessel performing an activity
together with its expected behavior on the travel segments Travel-Path Travel pathh comprising waypoints.
(i.e., the path between two consecutive waypoints). We <[location, <Map Each traveel segment points to a
characterize an agent’s behavior on a travell segment by its [speed, activity]>]> hashmap ofo speed to activity label
speed and the activity label (e.g., cruising, debbarking, docking, TABLE 3. SECONDARY CASE
E REPRESENTATION
or ramming). As shown in Figure 1, the Recognizer Feature Description
continuously observes the agent vessel’s actioons and compares Vessels-(Ref, The refeerence and approaching
them with the expected actions in all the acctive (still valid) Approaching) vessel tyypes
plans for consistency (see Table 1 for criteria)). It retains all the Relative-bearing Their reelative bearing
consistent plans and discards all inconsistent pplans. Primary-case-pointer Pointer to a primary case
Speed (Ref, Approaching) The speeds of the two vessels
TABLE 1. CONSISTENCY CHECK CRIT
TERIA
Avoidance-Actions Ordered d list of actions that the
Feature Value constrain nt referencce vessel will execute
Vessel-speed Within specifiedd threshold
Vessel-track Within specifiedd boundary Tables 2 and 3 display case repressentations for the PCB and
Bearing-to-waypoint Within specifiedd threshold SCB. The vessel type and its traveel path features, as defined
Bearing-changes Within specifiedd threshold above for our HTN plan recognizerr, are used to represent the
Distance-to-Waypoint Select next wayppoint PCB cases. In addition to the vesssel’s speed and its activity
label, the case-based algorithm also uses a Boolean feature because the former leverage add
ditional domain knowledge
(not shown in Table 2) to indicate whether thhe travel segment (i.e., expected paths).
is the last one of the entire path. 2. Compare the lead-time to recogn
nize a change in activity of
The SCB cases are represented by the two vvessel types, their the three algorithms. We conjecture
c that the plan
relative bearing, a pointer to the primary casse, the speeds of recognizers will outperform the behavior recognizer
the reference and the approaching vessels, annd the avoidance because they can detect morre subtle anomalies (e.g.,
actions that the reference vessel executes. Thee relative bearing moving to an unexpected dock).
feature measures how the two vessels’ bearinngs relate, and is
calculated as follows: B. Test Scenarios
rbx,y=rbx+rby We developed four simulated test scenarios
s that take place on
where rbx is the angle between the velocity vvector of vessel x a model of the Potomac River in Washington,
W DC. Besides the
and the vector pointing from vessel x to vesseel y. The value of real docks of National Harbor (NH H) and Old Town (OT), we
rbx,y ranges from 0 to 360. A value of 0 impliees the vessels are introduce a fictitious dock called Reeagan Airport Dock (RAD)
moving directly at each other and a value of 360 implies they to evaluate the impact of long-term m goals and intentions. We
are moving directly away from each other. developed the first two scenarios too examine the performance
on long-term intent following, wh hile the others allow us to
explore performance on short-teerm reactive events. The
descriptions of these scenarios (see Figure 3) are as follows:

1: NH

2: OT
O
3: RAD

Figure 2. Case-based plan recognition allgorithm

Figure 3. Scenario of two motorboats avoiding a head on collision
Plan recognition using the case bases involvees two concurrent
CBR tasks: waypoint intent recognitionn and collision 1. Routine Ferry (between NH to RAD): In this scenario the
avoidance recognition (see Figure 2). W Waypoint intent ferry travels between NH and RAD each hour. The paths
recognition proceeds as follows. Given thhe current state between the two docks are siimilar until they reach the
involving an agent vessel v, the best matchhing case cpcb is dock near OT, where they diverrge.
retrieved from the PCB and the criteria show wn in Table 1 are 2. Non-Routine Ferry (to OT from f RAD while ferrying
used to check consistency with cpcb. This is repeated until a between NH and RAD): This iss an anomalous counterpart
consistent case is found or the loop terminatess, whereupon v is to scenario #1. When the ferry y reaches RAD, instead of
marked as an anomalous situation (e.g., to rreport to a user). returning to NH, it ferries to OTT and back.
For the collision avoidance recognition taskk, if a case cscb 3. Collision Avoidance: This is a head-on collision
matching the current situation is found in thhe SCB, then the avoidance scenario among two o motorboats. They follow
waypoint recognition task is suspended annd the collision collision avoidance regulation ns, turning to starboard to
avoidance task becomes active. This latter tassk ends when v’s avoid a head-on collision when n they approach each other.
behavior becomes inconsistent with cscb’s exppected avoidance 4. Traffic Violation: This is an anomalous counterpart to
actions or it reaches the next waypoint in cpccb, at which time scenario #3, where one of the motorboats intentionally
control passes back to the waypoint intent recoognition task. disobeys collision regulations and continues on its path,
thereby simulating a pottentially dangerous and
IV. EVALUATION threatening situation to the otheer motorboat.
A. Objectives We encoded these scenarios in a maritime
m traffic simulation
environment, which we present next.
Our objectives were as follows:
1. Compare the activity labeling performancee of the two plan C. Test Environment for Maritimee Traffic Simulations
recognition algorithms against each otheer and versus a We modified a Navy simulation tool, named the Tactical
selected behavior recognition algorithm (see Section 4.D) Actions Officer (TAO) Sandbox, tot conduct our evaluation.
on simulated long-term and reactive sceenarios. Our two The TAO Sandbox is used to train Tactical
T Action Officers on
plan recognition algorithms should outperform the deployment and management off Navy assets for anti-
behavior recognition algorithm on the lonng-term activities submarine warfare and related missions [11]. We modified it
as follows: (1) its agents are controlled by our HTN algorithm, motorboat). Also, for dynamic replanning situations, like
and (2) it visually displays the behavior of our plan and HTN-PR, if the SCB retrieves a case cscb and the agent
behavior recognition algorithms in the Potomac River map. does not take any of cscb’s actions within a prespecified
First, we used an HTN planner, namely SHOP2PDDL+ [12] to time, then the activity is marked as anomalous.
generate the plans used by all the simulated agents controlling 3. MLN Behavior Recognizer (MLN-BR): This algorithm is
the vessels in our scenarios (i.e., motorboats and ferries). based on our recent work on probabilistic graphical
Among HTN planners, we selected SHOP2PDDL+ because it models [3], where we showed that MLNs performed well
can be used to generate behaviors in our continuous planning on maritime threat detection tasks. MLNs combine first-
domain. It uses a wait action to represent durative activities. order logic with a probabilistic interpretation to represent
For our scenarios, we encoded a set of SHOP2PDDL+ methods expert domain knowledge [13]. We used Alchemy [14] to
(e.g., for ferrying from one dock to another) that allows it to implement MLN-BR. It interfaces with the TAO Sandbox
generates plans such that, given a scenario’s start state and to obtain a representation of the world state as a set of
goal state, the resulting plan can be given to and executed by first-order logic clauses. MLN-BR performs MAP
the TAO Sandbox. To respond to a dynamic situation (e.g., inference for each agent to recognize its behavior. We
collision avoidance situations), we added an action that can be encoded knowledge for MLN-BR to identify the
used to perform conditional planning; it allows an agent to following activities: Debarking, Cruising, Docking, and
monitor the TAO Sandbox environment for any specified Docked. This knowledge includes five rules that encode
situations and, if matched, triggers a replan. The revised plan’s discretized values of a vessel’s speed, proximity to a
actions are then sent to the TAO Sandbox simulator for dock, and its bearing relative to docks. In addition, to
execution. handle the collision avoidance scenarios, we encoded
Second, we used Google Earth to accurately scale (from seven more logic rules. These rules involve additional
pixels to nautical miles) and depict the Potomac River area, features such as the possibility of head-on collision, and
including the NH, OT, and RAD ports (see Figure 3). We whether the agent is on its intended path to its next
modeled the waypoints for the four scenarios, their associated waypoint. Finally, the MLN is given a series of rules
ferry schedules, and boat transit speeds by observing actual regarding the vessel type (i.e., motorboat or ferry). For
maritime traffic on the Potomac River. We specified the example, the rule:
waypoints by considering the upstream and downstream traffic 5 !Speed(v,0,t) => !Action(Docked,v,t)
lanes that are marked with buoys. We added random
variations to the scenario instances to simulate realistic specifies that, if the speed of a vessel v is not 0 at time t,
maritime conditions and path variations. We included up to then v is not Docked at time t, and this rule has a weight
four concurrent agent-controlled vessels in a scenario to of 5. Here a weight is a measure of the strength of a
control for extraneous vessel interactions. Finally, we ran ten constraint for a given world. A world’s likelihood is lower
instances of each scenario and collected our observations. if it violates a constraint with a high weight.

D. Threat Detection Algorithms E. Metrics

We implemented two plan recognition algorithms for We used the following measures to compare the algorithms:
comparison with the best-performing behavior recognition 1. Activity Labeling Accuracy: This is the number of
algorithm from our previous studies [3]. correctly predicted labels for an event type divided by the
total number of that type of event. Our scenarios included
1. HTN Plan Recognizer (HTN-PR): This implements the the following event labels: Avoiding Collision, Cruising,
HTN plan recognition algorithm described in Section 3. Debarking, Docked, Docking, and Anomaly. We
The Recognizer is connected to and observes the TAO computed the accuracy for each label type and the
Sandbox environment (e.g., the vessels’ activities). It uses accuracy over all the label types.
the same planning model given to SHOP2PDDL+ (see
Section 4.C). Also, for each of the four scenarios, we 2. Anomaly Labeling Latency: Each time step during the
provide this planner with the set of possible goal states course of an activity represents an event, which the
(e.g., ferry passengers to a particular dock). algorithms label. The ability to label an anomalous
For dynamic situations, the Recognizer periodically activity as soon as it starts is desirable for the algorithms.
replans to mimic the conditional replans that the agent Therefore, we measure the time elapsed between the onset
might execute. It uses these to detect anomalies by of an anomaly and the time it takes for an algorithm to
noticing when the actual and expected actions deviate correctly detect it. Smaller time lags imply a better
during a specified time frame (e.g., as in scenario #4, performing algorithm.
where the agent-controlled motorboat deliberately F. Results
disregards collision avoidance regulations). Table 4 displays the accuracy of three algorithms on our four
2. CB Plan Recognizer (CB-PR): This implements the case- scenarios. In Scenarios 1 and 2, which pertain to longer-term
based plan recognition algorithm described in Section 3. behaviors, the plan recognition algorithms significantly
We manually populated the PCB with 4 cases (one per outperform MLN-BR. For instance, in the Routine Ferry
vessel instance) and the SCB with 2 cases (one per Scenario (#1), HTN-PR and CB-PR recorded higher
accuracies than MLN-BR (0.859, 0.839 vs. 0.776). However, TABLE 5. ANOMALY LABELING LATENCY (SECONDS)
the performance difference among the two plan recognition Scenario #2 #4
algorithms is small. In Scenario #2, concerning non-routine HTN-PR 118.0 90.4
ferries, HTN-PR and CB-PR again significantly outperform CB-PR 27.0 64.7
MLN-BR (0.928, 0.938 vs. 0.564), which cannot detect MLN-BR * 10.7
anomalies because it has no knowledge of the intended routes.
Furthermore, CB-PR outperforms HTN-PR overall and on V. DISCUSSION
anomaly detection for this scenario (0.984 vs. 0.931). Many methods have been investigated for plan and behavior
Performance on the reactive scenarios is mixed. In recognition (e.g., [15][16][17]). They differ along many
particular, in Scenario #3, one of the plan recognition dimensions, such as their agent relation (e.g., keyhole,
algorithms (CB-PR) significantly outperforms the behavior intention, adversarial), what they infer (e.g., action, plan,
recognizer (MLN-PR) on the avoidance activity (1.0 vs. goal), model representation type (e.g., decision theoretic,
0.773). However, in Scenario #4, MLN-BR outperforms both probabilistic), whether they are generative or discriminative,
the plan recognition algorithms on anomaly detection (0.881, their model family (e.g., Bayesian, MDP, random field), and
0.905 vs. 0.971), although the differences were not significant. how they detect anomalies. These have been applied to a
We believe this occurred because MLN rules generalize better plethora of tasks, such as adversarial strategy detection, user
in a reactive situation, a conjecture that we will explore in our modeling, human activity recognition, and threat detection,
future work. CB-PR outperforms HTN-PR in Scenario 4, and which is our focus. However, little published research exists
predicts all avoidance actions correctly on Scenario 3. on applying plan recognition techniques for maritime threat
TABLE 4 . AVERAGE ACCURACY PER ACTIVITY (10 RUNS) recognition, although Nicolescu and her colleagues have
studied this topic in recent years, including the application of
hidden Markov models (HMMs) and spreading activation
techniques for detecting potentially deceptive behavior [18].
Our own related work has most recently focused on
applying probabilistic graphical models to synthetic and real-
world threat recognition scenarios [3]. We found that Markov
logic networks possess some advantages in comparison with
HMMs and conditional random fields, though they require
substantial care to design and time to train. Our experience
with these and simpler approaches for threat recognition [19]
led us to consider knowledge-intensive plan recognition
approaches that can provide additional constraints to the
prediction task. From our prior work and this investigation,
our observation is that plan recognition algorithms can be
preferable to behavior recognition algorithms in scenarios with
longer-term anomalies that are detectable using background
knowledge, though the inverse can be true for reactive
scenarios. This suggests that a hybrid approach may be useful
for maritime threat detection, which we will investigate as part
Therefore, this supports our first conjecture that these plan of our future research.
recognizers can recognize some longer-term behaviors more
accurately than this behavior recognition algorithm. VI. CONCLUSION
Table 5 displays average anomaly labeling latencies for the Existing algorithms for anomaly and threat detection
anomaly recognition scenarios (i.e., #2, and #4). For longer- predominantly use statistical normalcy models. Although these
term intent violations (Scenario 2), CB-PR substantially are somewhat successful in thwarting short-term reactive
outperforms HTN-PR (27 vs. 118 seconds). Thus, CB-PR is a changes, they cannot detect subtle but critical violations in
more accurate activity labeler and is also faster at detecting an routine plans. To address this issue, we developed plan
anomalous situation. In contrast, MLN-BR’s time is infinite recognition algorithms to detect longer-term and reactive
since it does not detect any anomaly in this scenario. violations of expected vessel behaviors in a simulated
However, it performs much better in the reactive Scenario 4. maritime environment. In particular, we developed two
For instance, it detects collision avoidance anomalies after an consistency-based plan recognition algorithms based on HTN
average of 10.7 seconds after its onset compared to at least a and case-based techniques. We evaluated them on a set of four
minute (i.e., 64.7 seconds) for the other algorithms. scenarios and found evidence that they significantly
outperform a behavior recognition algorithm on longer-term
plan recognition tasks.
To our knowledge, this is the first published evaluation of
plan recognition algorithms for maritime threat detection
tasks. This is an initial study; further research is needed to [14] Alchemy (2011). Alchemy ─ Open source AI.
[http://alchemy.cs.washington.edu]
better understand the pros and cons of these algorithms. For
[15] Kautz, H., & Allen, J.F. (1986). Generalized plan recognition.
example, our evaluation was limited in the types of scenarios Proceedings of the Fifth National Conference on Artificial Intelligence
and the number of concurrent agents active in a scenario. In (pp. 32–38). Philadelphia, PA: AAAI Press.
our future work, we will extend our evaluation to scenarios of [16] Ramirez, M., & Geffner, H. (2009). Plan recognition as planning.
Proceedings Twenty First International Conference on Artificial
coordinated small boat attacks, denser traffic conditions, and
Intelligence (pp.1778-1783). Pasadena, CA: AAAI Press.
schedule violations to investigate the generality of our [17] Goldman, R.P., Geib, C.W., Kautz, H., & Asfour, T. (Eds.) (2011). Plan
approach. Also, although we considered stochastic variations recognition (Report from Dagstuhl Seminar 11141).
in our scenarios, they were limited and our consistency-based [drops.dagstuhl.de/opus/volltexte/2011/3195/pdf/dagrep_v001_i004_p0
01_s11141.pdf]
approaches were not stochastic. Therefore, we will examine
[18] Nicolescu, M., Nicolescu, M., & Louis, S. (2012). Context-based intent
probabilistic approaches to plan recognition in our future understanding for autonomous systems in naval and collaborative
research. Finally, we assumed that a library of HTN plans is robotics applications. Presentation at the ONR Cognitive Science and
available to our system at the outset. However, in practice, Human Robot Interaction Program Review, June 2012.
[19] Auslander, B., Gupta, K.M., & Aha, D.W. (2011). Comparative
they will need to be acquired from subject matter experts or
evaluation of anomaly detection algorithms for local maritime video
learned from observations. We plan to explore other surveillance. In Proceedings of the Sensors, and Command, Control,
hierarchical and case-based learning algorithms to Communications, and Intelligence (C3I) Technologies for Homeland
automatically acquire such models (e.g., [20][21]). Security and Homeland Defense X. Orlando, FL: SPIE.
[20] Geib, C.W., & Goldman, R.P. (2009). A probabilistic plan recognition
ACKNOWLEDGMENT algorithm based on plan tree grammars. Artificial Intelligence, 173(11),
1101-1132.
We thank the Office of Naval Research for funding this [21] Kerkez, B. & Cox, M T 2003. Incremental case-based plan recognition
research. with local predictions. International Journal on Artificial Intelligence
Tools: Architectures, languages, algorithms, 12(4), 413-464.
REFERENCES
[1] Lipton, A.J., Heartwell, C.H., Haering, N., & Madden, D. (2002).
Critical asset protection, perimeter monitoring, and threat detection
using automated video surveillance. In Proceedings of the Thirty-Sixth
Annual International Carnahan Conference on Security Technology.
Atlantic City, NJ: IEEE Press.
[2] RemoteReality (2011). OmniAlert360. Retrieved on 20 November 2011
from [http://www.remotereality.com].
[3] Auslander, B., Gupta, K.M., & Aha, D.W. (2012). Maritime threat
detection using probabilistic graphical models. In Proceedings of the
Twenty-Fifth International FLAIRS Conference. Marco Island, FL:
AAAI Press.
[4] Cook, D.J., Augusto, J.C., & Jakkula, V.R. (2009). Ambient
intelligence: Technologies, applications, and opportunities. Pervasive
and Mobile Computing, 5, 277-298.
[5] Jarvis, P.A., Lunt, T.F., & Myers, K.L. (2004). Identifying terrorist
activity with AI recognition technology. Proceedings of the Nineteenth
National Conference on Artificial Intelligence (pp. 858-863). San Jose,
CA: AAAI Press.
[6] Genesereth, M.R. (1984). The use of design descriptions in automated
diagnosis. Artificial Intelligence, 24, 411-436.
[7] Schmidt, C.F., Sridharan, N.S., & Goodson, J.L. (1979). The plan
recognition problem: An intersection of artificial intelligence and
psychology, Artificial Intelligence, 10(1), 45-83.
[8] Kautz, H. (1987). A formal theory of plan recognition. Doctoral
Dissertation: Department of Computer Science, University of Rochester,
NY.
[9] Nau, D., Au, T.-C., Ilghami, O., Kuter, U., Murdock, J.W., Wu, D., &
Yaman, F. (2003). SHOP2: An HTN planning system. Journal of
Artificial Intelligence Research, 20, 379-404.
[10] Aamodt, A., & Plaza, E. (1994). Case-based reasoning: Foundational
issues, methodological variations, and system approaches. AI
Communications, 7, 39-59.
[11] Auslander, B., Molineaux, M., Aha, D.W., Munro, A., & Pizzini, Q.
(2009). Towards research on goal reasoning with the TAO Sandbox.
(Technical Note AIC-09-155). Washington, DC: Naval Research
Laboratory, Navy Center for Applied Research on Artificial Intelligence.
[12] Molineaux, M., Klenk, M., & Aha, D.W. (2010). Planning in dynamic
environments: Extending HTNs with nonlinear continuous effects. In
Proceedings of the Twenty-Fourth AAAI Conference on Artificial
Intelligence. Atlanta, GA: AAAI Press.
[13] Domingos, P., & Lowd, D. (2009). Markov logic: An interface layer for
AI. Morgan & Claypool.