Title: A Multi-Agent System for Tracking the Intent of Surface Contacts in Ports and Waterways
1A Multi-Agent System for Tracking the Intent of
Surface Contacts in Ports and Waterways
- Tan, Kok Soon Oliver
- Project Manager
- C4IT-IKC2
- DSTA
- tkoksoon_at_dsta.gov.sg
2Agenda
- Introduction
- Concepts
- Multi-agent System Design
- System Validation Scenarios
- Recommendations and Conclusion
3Introduction
- A thesis on modeling the intent of surface
contacts with a multi-agent system (MAS) for
asymmetric threat identification in busy ports
and waterways - Inspired by similar work done in the area of air
threat assessment in Air Defense Laboratory (ADL)
Ozkan 2004, NPS
4Thesis Questions
- How can surface contact intent be modeled with a
MAS for the identification of potentially hostile
behaviors and threats in ports and waterways? - Will the models be sufficiently realistic to be
used as a decision aid in maritime security?
5Why a MAS?
- A multi-agent model is a distributed
intelligence model that is a natural solution
of a large-scale real-world problem1 - The real world problem is physically distributed
- Every surface contact is an autonomous entity
that we are interested in knowing its probable
intentions - The knowledge to solving the real world problem
is widely distributed and heterogeneous - No one agent or system is "knowledgeable" enough
to trawl and mine databases, process real-time
sensor data, monitor for rule violations or
suspicious behaviours etc - The sources of data are distributed over networks
- Naturally this encourages us to take a
distributed view of a solution for the real world
problem and - The real world problem is too complex to be
analysed as a whole - There are too many parameters and constraints to
be considered altogether. Local approaches,
partitioning the large problem into smaller and
more tractable sub-problems, can produce results
quickly.
1. Ferber, J., Multi-Agent Systems An
Introduction to Distributed Artificial
Intelligence, Addison-Wesley, 1999.
6MAS Objectives
- A multi-agent system (MAS)
- To help the human operator sieve through hundreds
of surface contacts - To integrate intelligence and information from as
many sources as possible - To highlight any suspicious or potentially
hostile surface contacts
7Requirements of the MAS
- Support rules and regulations of a Vessel Traffic
Service (VTS) such as - Traffic Separation Scheme (TSS), part of the
International Navigation Rules defined by the
International Maritime Organization - the 1972 Collision Regulations (72 COLREGS)
- International Ships and Port Facilities Security
(ISPS) Code (To be implemented) - all other practices of safe navigation and
prudent seamanship - Predefined safe speed limits in TSS
- Safe speed limits for different surface track
types
8Requirements of the MAS cont
- Use Surface Warfare Threat Assessment cues and
corresponding perception of threat Liebhaber
2002, SPAWAR Systems Center, San Diego - Obtained through empirical and observational
studies of the threat assessment process by
experienced surface warfare officers - Each cue has a Threat Level Change Rating (TCR)
that changes the threat level posed by a surface
contact
1. Platform/Weapon Envelope/ESM
2. Origin-Flag
3. Range/Distance from Own-Ship (subsumed under CPA)
4. Heading (subsumed under CPA)
5. Closest Point of Approach (CPA)/Speed
6. Number of vessels (To be implemented)
7. Own support in area (To be implemented)
8. Destination
9. History/Voice communication
10 Sea Lane/Other intelligence
11. Superstructure Type (To be implemented)
12. Coordinated Activity (To be implemented)
9Requirements of the MAS cont
- Use information from ship-borne
- Automatic Identification system (AIS)
- Transponder for large vessels (gt300gt)
- International Maritime Organization (IMO)
recommendation - Harbor Craft Transponder system (HARTS)
- For smaller vessels
- Applies to the Port of Singapore only
1. Track Type
2. Callsign (To be implemented)
3. IMO Number (Lloyds Register Number) (To be implemented)
4. Maritime Mobile Service Number (To be implemented)
5. ETA (To be implemented)
6. Destination (To be implemented)
10Thesis Scope
- Identify and track the intent of surface contacts
- Borrowing the ideas and techniques suggested for
identifying air threats in the Air Defense
Laboratory (ADL) and use them to identify
asymmetric maritime threats - The thesis does not cover the issue of track
detection i.e. assumes perfect instantaneous
detection with 100 reliability - The issue of interdiction when a potentially
hostile track has been identified is also beyond
the scope of this thesis
11Some Concepts
- Traffic Separation Scheme (TSS)
- Security Zones for HVUs
- Security Zones for Restricted Areas
- Areas-To-Be-Avoided (ATBA)
- Safe Speed Limits
Skip Concepts
12Traffic Separation Scheme (TSS)
- A TSS is a sea lane with a predefined traffic
direction - A TSS may also has a predefined safe speed (for
prudent seamanship) - A violation occurs when a track is traveling
against traffic direction or is traveling at an
excessive speed
13Security Zones for HVUs
- Every High Value Unit (e.g. cruise liner, tanker)
have their own predefined multiple security zones - Only some type of tracks (e.g. Police Coast
Guards) are allowed within these security zones - Each security zone is defined with an alert time
threshold (represents a measure of urgency when
these zones have been infringed)
14Security Zone Violation Example
- A security zone violation occurs i.e. a track is
coming in too near, too soon, if an unauthorized
track has - a CPA (Closest Point of Approach) within a zone,
and - a TCPA (Time to CPA) below alert time threshold
-
Too near! Too soon!
CPA
TCPA 3min
Radius 0.2nm, Alert Time 15min
Radius 0.5nm, Alert Time 10min
Radius 0.8nm, Alert Time 5min
15Security Zones for Restricted Areas (Static HVUs)
- Restricted areas (e.g. harbor, oil refineries,
military installations) can have their own
predefined multiple security zones - Only some type of tracks (e.g. Police Coast
Guards) are allowed within security zones - Each security zone is defined with an alert time
threshold
Radius 0.2nm, Alert Time 15min
Radius 0.5nm, Alert Time 10min
Radius 0.8nm, Alert Time 5min
16Areas-To-Be-Avoided (ATBA)
- Restricted areas (e.g. harbor, oil refineries,
military installations) - Only allow certain types of tracks (e.g. Police
Coast Guards) or certain types of track activity
within these areas - An ATBA violation occurs when an unauthorized
track intrudes into a restricted area
17Safe Speed Limits
- Some locations or restricted areas (e.g. harbor)
may only allow tracks to travel at predefined
speed limits - Speed limits can be defined for different track
types - A violation occurs when a track exceeds any of
these speed limits
18The Compound Multi-agent System
- A compound multi-agent system (MAS) designed for
surface contact intent tracking - Each surface contact is represented by a track
agent - Every track agent has a nested MAS (Russian
Doll)
19Anatomy of a Track Agent
Friendly Intent Agent
Neutral Intent Agent
Potentially Hostile Intent Agent
Unknown Intent Agent
Composite Agents
ATBA Zone Track Activity Violation Blend
ATBA Zone Track Type Violation Blend
Security Zone Violation Blends
Speed Threshold Violation Blend
Speed Violation Blend
TSS Heading Violation Blend
Security Zone Violation Blends
Cognitive Agents
Speed Threshold Violation Agent
TSS Heading Violation Agent
Speed Violation Agent
Location Agent
Area-To-Be-Avoided (ATBA) Violation Agent
Security Zone Violation Agent
Reactive Agents
Track Flag Data Ticket
Track Origin Data Ticket
Track Destination Data Ticket
Track ESM Data Ticket
Track Type Data Ticket
Track Position Data Ticket
Track Activity Data Ticket
Track Comm Data Ticket
Track Heading Data Ticket
Track Speed Data Ticket
20The Compound Multi-agent System cont
- Agents in the nested MASs continuously process
incoming information about their respective
surface contacts - Agents communicate and co-ordinate in order to
discover the likely intent of surface contacts
21Conceptual Blending
- Conceptual Blending1 is a theory about how humans
process the information coming from the
environment and how humans rationalize the events
happening around them - Blending is a set of mental operations for
combining cognitive models in a network of mental
spaces - Mental spaces are small conceptual packets
1. Gilles, F., Turner, M., The Way We Think,
Basic Books, New York, 2002
22Conceptual Blending cont
- Mental spaces are connected to long-term
schematic knowledge called frames e.g. - The frame of sailing along a ferry route, or
- The frame of traveling inside a maritime traffic
separation scheme (TSS), - Long-term specific knowledge such as a memory of
an event such as past track incursions into
Area-To-Be-Avoided (ATBA) zones. - Mental spaces are interconnected in working
memory which can be modified dynamically
23Conceptual Blending cont
Generic Space
A Basic Conceptual Integration Network
Input Space 1
Input Space 2
Blend
- Building a conceptual integration network
involves setting up several mental spaces. - Two input mental spaces with cross-space mapping
to connect counterparts in these input mental
spaces - However not all elements and relations from the
input spaces are projected into the blend. - Generic spaces are used for the generic
structures they contain to guide the selective
projection of elements from the input spaces into
blended spaces - The blended space is the mental space where,
during blending, the structure from the input
mental spaces is projected onto, represented by
the dotted lines
24Conceptual Blending cont
- Any mental space can participate in multiple
networks. - Complex integration networks can be built with
arrays of mental spaces that are connected
through blending operations.
25Conceptual Blending Examples
Generic Space
CPA lt Security Zone Radius
TCPA lt Security Zone Alert Time
Track Type ? Allowed Track Types
Track Type
Allowed Track Types
Identity Vital Relation
Security Zone Radius
Distance Vital Relation
Track CPA (Closest Point of Approach)
Time Vital Relation
Security Zone Alert Time
High Value Unit
Track TCPA (Time to CPA)
Track
Blend
Security Zone Violation
- Example of how a Security Zone violation is
detected
26Conceptual Blending Examples
Generic Space
Track Activity ? Allowed Activity Type
Track Location Zone Name
Track Activity
Allowed Activity Type
Activity Vital Relation
Activity Vital Relation
Location Vital Relation
Track Location
Zone Name
ATBA Zone
Track
Blend
ATBA Zone Track Activity Violation
- Example of how a ATBA Zone Track Activity
violation is detected
27The CMAS Library
- The communication and coordination among many
different agents in the nested MAS is achieved
using the Connector-based Multi-agent Simulation
Library (CMAS) John Hiles, NPS - The basic elements for agent communication and
control within the CMAS framework are connectors. - The agents use these connectors to externalize
portions of their internal states into the
multi-agent environment. - Connectors are like plugs and receptacles that
can be extended or retracted - Signaling and coordination between the two agents
occur when there are matching pairs of
plug-receptacle connectors and the connectors get
connected - Stimergy (communication through change of local
environment) among agents
Agent 2
Extended response connector (Receptacle)
Plug-Receptacle match
Extended stimulus connector (Plug)
Agent 1
Retracted connector
28A MAS of MASs (Russian Doll)
- A track agent appears as a single agent that
exists in another external MAS environment - In this external MAS environment, there is a
layer of regional agents that monitor behaviors
of all track agents - Two types of regional agents detect coordinated
behavior that resembles an impending swarm or a
wolf-pack attack
29Detection of Coordinated (Swarm/ Wolf-pack)
Attack on a moving HVU
Too near! Too soon! Too many!
- If two or more track have
- CPAs to a HVU (High Value Unit) that are very
close e.g. 0.1 nm, and - TCPAs violations against the same HVU that are
about to occur within a very short period of time
e.g. 5 mins
The MAS will consider multiple near-simultaneous
security zone violations a possible sign of an
impending coordinated attack i.e. too near, too
soon, too many Note A wolf-pack attack is a
common maritime terrorist attack tactic
comprising of a cluster of small terrorist craft
approaching and surrounding a larger target craft
from multiple directions simultaneously
30Detection of Coordinated (Swarm/Wolf-pack) Attack
on a static HVU
Too near! Too soon! Too many!
- If two or more track have
- CPAs to a restricted location (static HVU) that
are very close e.g. 0.1 nm, and - TCPAs violations against the same location that
are about to occur within a very short period of
time e.g. 5 mins
The MAS will consider this a possible sign of an
impending coordinated attack i.e. Too near, Too
soon, Too many
31Anatomy of a Regional Agent
32Conceptual Blending Examples
Too much of a coincidence?
Generic Space
HVU(A) HVU(B)
(CPA(A) CPA(B)) lt CPA_DIFFERENCE_THRESHOLD
(TCPA(A) TCPA(B)) lt TCPA_DIFFERENCE_THRESHOLD
HVU (A)
HVU (B)
Identity Vital Relation
Distance Vital Relation
CPA(B)
CPA(A)
Time Vital Relation
Track B
Track A
TCPA(B)
Security Zone Violation Blend B
Security Zone Violation Blend A
TCPA(A)
Blend
Swarm Detection Blend
- Example of how a Coordinated Attack
(Swarm/Wolf-pack) by 2 or more different tracks
on the same HVU is detected by a Regional Agent
33The Intent Agent
- The top layer of agents of the nested MAS
environment inside a track agent - Each intent agent has a corresponding intent
model - Four intent agents
- Friendly,
- Neutral,
- Potentially Hostile, and
- Unknown
- Intent agents use information provided by
internal agents from the lower layers as well as
from external regional agents
34Anatomy of an Intent Agent
MARSEC Level (bias)
Weighting Strategy
Swarm Detection (Track)
Swarm Detection (Location)
Weighting Agents
ATBA Zone Track Activity Violation
ATBA Zone Track Type Violation
Security Zone Violation
Speed Threshold Violation
Speed Violation
TSS Heading Violation
Track Type
Track Flag
Track Origin
Track Destination
Track Comm
Track ESM
35Competitive Intent Models
- An Intent agent is a composite agent
- Family of weighting agents is responsible for
obtaining information - User-defined weights (similar to Threat Level
Change Ratings) assigned to each piece of track
information (attributes and violations) - The intent model in an intent agent is
represented by a weighting strategy - Weighting agents receive track information on
track attributes and track violations and informs
the weighting strategy - Weighting strategy computes a weighted score
using a set of user-defined weights - The intent models will compete and the one with
the highest score represents the current intent
of the track
36Weighting Biases based on Regional Intelligence
- Maritime Security (MARSEC) Levels
- Warning against unidentified potential threats
- Equivalent to HSAS
- Heightens/Lowers the alertness of the weighting
strategies by applying biases to the computed
weighted scores.
37The VTS-C2 MAS System
38Features of the VTS-C2 system
- A Java-based mock C2 (Command Control) system
- Supports geo-rectified maps, tactical overlays
and symbol drawing, graphical and tabular
displays of C2 information - Shows graphics representing tracks, TSSes, and
restricted areas - Integrated CMAS-based (Connector-based
Multi-agent Simulation Library) compound MAS - Integrated Simkit-based DES (Discrete Event
Simulation) simulator Arnold Buss, NPS - Tracks are linear uniform movers with delays at
waypoints - Proximity sensors are used to report location of
tracks
39Capabilities of the VTS-C2 MAS
- Ability to detect future incursions into the
security zones of HVU (high value units) - Ability to detect future incursions into
restricted areas e.g. cruise center, oil
refineries, military installations - Ability to detect illegal activities in
restricted areas e.g. fishing in non-fishing zone - Ability to detect TSS (traffic separation
schemes) violations e.g. against traffic
direction, stopping in TSS termination zones - Ability to detect speed violations in restricted
areas e.g. harbor - Ability to detect atypical track behaviors e.g.
excessive speed
40Capabilities of the VTS-C2 MAS cont
- Ability to perform surface threat assessment
based on tracks attributes e.g. platform, flag,
origin, ESM, destination - Ability to detect VTS (Vessel Traffic Service )
violations e.g. collision detection,
wrong/unknown destination, no verbal
communication - Ability to detect coordinated maneuvers/attacks
e.g. swarm/wolf-pack - Ability to incorporate regional intelligence e.g.
MARSEC levels
41System Architecture
Databases (Lloyds, ICA)
Pre-defined Information
Java-based VTS-C2 system
Hourly/Ad-hoc Reports (Police Coast Guards/
Military Patrols)
MARSEC Level
TSS Definitions (traffic direction, speed limits)
Anecdotal Anomalies Detection
Compound MAS
24-hour Offshore Advance Reports (International
Maritime Organisation Standard Ship Reporting
System)
MAS of MASs
Security Zone Definitions (CPA radius and alert
time)
Operational Anomalies Detection
Ship Manifests (Cargo/ Crew/ Passenger
information) (ICA)
ATBA (Area-To-Be-Avoided) Definitions (allowed
track types, allowed track activity)
Information Sources (MPA)
Track Type, Callsign, IMO Number (Lloyds
Register Number), Maritime Mobile Service Number,
ETA, Destination
Safe Speed Limits for each track type
Track Position, Speed, Heading, Destination
Automatic Identification System
Maritime Sensors (Simkit-based Discrete
Event Simulator)
Harbor Craft Transponder System
Safe Speed Limits for certain locations and zones
To be implemented
42Pre-defined Information Settings
43Weight and Bias Settings
44Agent Threshold Parameters
45Intent Scores Information
46Validation Process
- Four validation sessions held with four groups of
surface warfare assessment experts or naval
officers from the Republic of Singapore Navy
(RSN) and the US Navy - Participants have more than 100 years of harbor
security, patrol or at-sea experience between
them - Participants are first briefed on the features of
the MAS and the mock VTS-C2 system - Participants are next presented with several
discrete-event simulations on scenarios involving
the Port of Singapore and the surrounding
waterways
47Validation Process cont
- Each scenario features multiple surface contacts
of different types, moving in an area that is
populated with traffic separation schemes and
restricted areas - The scenarios will feature different kinds of
hostilities that may exist but the participants
are not told of the details in advance
48VTS-C2 System Demo
Skip Scenarios
49Sample Scenario 1(TSS violations, Impending
collision)
TSS violation (speed and heading) and an
impending collision between a leisure craft and a
cruise liner
50Sample Scenario 2 (Coordinated attacks by
multiple tracks)
Possible coordinated attack by two fishing
vessels on SZone3
51Sample Scenario 3 (Incursion into security zone
around HVU)
Detected incursion by fishing vessel in the
security zone around tanker
52Validation Results
- Very encouraging responses from the participants
- good proof of concept that demonstrates how a
decision support tool can help the decision maker
identify potentially hostile contacts - Officers from the RSN commented that the MAS can
be an important decision support tool in their
existing C2 systems
53Validation Results cont
- Some concerns
- Although the system is able to process large
amount of information, there may still be an
overwhelming information glut - Intent labels not semantically suitable according
to operational doctrine if the MAS was to be
integrated into an existing C2 system - False alarms that may arise due to the heavy
traffic conditions in the Port of Singapore
compounded by clutter caused by non-moving
surface contacts Need to select weights
carefully to reduce the number of false alarms
False alarms is better than no alarms - System is highly dependent on accuracy and
reliability of information sources e.g. sensors,
humans etc.
54Future Work To Be Done
- Need to fine-tune the MAS and verify that system
works well against real world vessel traffic
situations in the waters of Singapore. - The system may be tested during maritime security
experiments - Further validation with objective measures of
performance - Type I (false negatives) and Type II (false
positives) errors, - Number of factors that the system can process as
compared to a human operator, - Time taken by the system to identify hostilities
as compared to a human operator i.e. amount of
lead time the system is able to provide in
situations of hostilities
55Future Work To Be Done cont
- Detect more unusual track maneuvers
- Many maneuvers / zig-zags
- Suspicious course changes that seem to match the
movement of a HVU - Monitor course/heading of tracks in more detail
(e.g. in terms of Steady and closing/opening or
Turn to closing/opening) - Hiding or evading from PCG/Military Patrols
- More co-ordinated activities among tracks e.g.
Simultaneous attacks on multiple HVUs or
restricted areas - Additional VTS violations
- Failure to submit Offshore Advance reports, and
- Wrong/unknown destinations
- Incorporate specific intelligence based on
- track attributes e.g. track type, origin,
activity and - historical data e.g. piracy reports
56Future Work To Be Done cont
- Beyond more than just rules, it is also possible
to have complex cognitive agents that can learn
and adapt - automatically learn appropriate weight settings
to reduce false alarms, or - automatically retire agents that are producing
too many false alarms, or - Automatically re-adjust security zone radii
according to traffic conditions, or - have the ability to forgive, over time, tracks
for their past violations - Act as proxies to external entity/relationship
engines, information fusion/search engines or web
services i.e. distributed intelligence - Pro-active search by agents for anecdotal
anomalies i.e. form a paper trail from
information sources such as ship registries, sail
plans, Offshore Advance reports, cargo/passenger
manifests etc
57Conclusion
- Thesis question 1 How can surface contact
intent be modeled with a MAS for the
identification of potentially hostile behaviors
and potential threats in ports and waterways? - A multi-agent system has been developed to track
the intent of multiple surface contacts moving in
ports and waterways. - Four intent models have been developed based on
VTS rules, surface warfare threat assessment cues
and track behaviors - Thesis question 2 Will the models be
sufficiently realistic to be used as a decision
aid in maritime security? - Face validation showed that the system can be a
useful decision support tool in the maritime
security of Singapore