BRECCIA: A Multi-Agent Data Fusion and Decision …...BRECCIA Agent P.L. Logic Module GeoWave...

BRECCIA: A Multi-Agent Data

Fusion and Decision Support

Framework for Dynamic Mission

Planning

DDDAS 20171

David Sacharny, Tom Henderson, Robert

Simmons, Amar Mitiche, Xiuyi Fan and

Taylor Welker

DDDAS 2017

Cambridge, MA

7 August 2017

Colleagues

Robert

Simmons

Amar Mitiche

INRS

Montreal

Xiuyi Fan

Nanyang

Technological

University

Tom HendersonDavid

Sacharny

Taylor

Welker

DDDAS 20172

Acknowledgment

This material is based upon work

supported by the Air Force Office of

Scientific Research under award

number FA9550-17-1-0077

(DDDAS-based Geospatial Intelligence)

DDDAS 20173

DDDAS 20174

BRECCIA

BRECCIA

DDDAS 20175

BRECCIA and DDDAS

DDDAS 20176

1. Applications Modeling: e.g., Wind/Obscurant Simulations

BRECCIA and DDDAS

DDDAS 20177

2. Advances in Mathematical and Statistical Algorithms:

e.g., Probabilistic Logic

No Weak

Methods

BRECCIA and DDDAS

DDDAS 20178

3. Application Measurement Systems and Methods:

e.g., Path Planning

BRECCIA and DDDAS

DDDAS 20179

4. Software Infrastructures and System:

e.g., BRECCIA Multi-agent Server

BRECCIA: Summary

• Provides middleware for:

• real-time coupling of computation and

knowledge

• across heterogeneous platforms

DDDAS 201710

BRECCIA: Summary

• Provides uncertainty analysis for

• mission planning

• involving combination of:

• human statements

• simulation results

• sensor measurements

DDDAS 201711

BRECCIA: Summary

• Agents driven by uncertainty reduction:

• identification of major uncertainty sources

• uncertainty quantification

• propose measures for uncertainty reduction

DDDAS 201712

Two Main Results

• Probabilistic Logic

• New approach:

• system of nonlinear equations

• Results on large systems

• Multi-agent Middleware

• Real-time agent mission assessment

and replanning

DDDAS 201713

Uncertainty in Knowledge

BasesGiven:

• A set of propositions

(e.g., 𝑆1: 𝑈𝐴𝑉1 is operational)

• A set of probabilities for the propositions

(e.g., (𝑆1)[0.9])

Then, given a query (e.g., Q: Mission_OK),

determine its probability

DDDAS 201714

Probabilistic Logic

• SAT (Satisfiability Problem)

Given a propositional calculus formula,

find a truth assignment to each logical

variable so that the formula is true

E.g., S ≡ P & (PQ) (i.e., Modus Ponens premises)

P True and Q True satisfies S

DDDAS 201715

Probabilistic Logic

• PSAT (Probabilistic SAT Problem)

(Simple version!) Given a CNF formula,

and a probability assignment for each

conjunct, find a consistent probability

assignment for a query formula

E.g.: [0.7] 𝑆1: P

[0.7] 𝑆2: ¬P v Q

[?] Query: Q

DDDAS 201716

To Solve: Count Models

DDDAS 201717

% Counts

𝑡1𝑡2𝑡3𝑡4

Models

P Q

0 0

0 1

1 0

1 1

Sentences

P

¬P v Q

Probabilities

0.7

0.7

ω∈Ω models P 𝑡3+𝑡4 = 0.7


DDDAS 201718

Counts


Models

P Q

0 0

0 1

1 0

1 1

Sentences

P

¬P v Q

Probabilities

0.7

0.7

ω∈Ω models ¬P v Q 𝑡1+ 𝑡2 + 𝑡4 = 0.7


DDDAS 201719

Solve:

𝑡3 + 𝑡4 = 0.7

𝑡1+ 𝑡2 + 𝑡4 = 0.7

𝑡1+ 𝑡2+ 𝑡3+ 𝑡4 = 1 add constraint

So solve:

0 0 1 11 1 0 11 1 1 1


= 0.70.71

E.g.:


=

00.30.30.4


DDDAS 201720

Solve:

𝑡3 + 𝑡4 = 0.7

𝑡1+ 𝑡2 + 𝑡4 = 0.7

𝑡1+ 𝑡2+ 𝑡3+ 𝑡4 = 1 add constraint

So solve:

0 0 1 11 1 0 11 1 1 1


= 0.70.71

E.g.:

𝒕𝟏𝒕𝟐𝒕𝟑𝒕𝟒

=

0.30.00.30.4


DDDAS 201721

So solve query, e.g., P(Q), sum probabilities of models

of formula:

E.g., for


=

0.30.00.30.4

then P(Q) = 𝒕𝟐+ 𝒕𝟒 = 0.4

E.g., for


=

0.00.30.30.4

then P(Q) = 𝒕𝟐+ 𝒕𝟒 = 0.7

Thimm’s Formulation

DDDAS 201722

A constraint r, is a disjunction with a probability.

Ω is the set of all complete conjunctions (a literal from

every logical variable appears once and only once in a

complete conjunction).

Geometric View of Query

DDDAS 201723

Example of Ω

DDDAS 201724

Linear Formulation Issues

• Exponential Complexity (in number of

sentences)

• Uses SAT solvers to produce the matrix

• Solver Complexity:

• Constraints: 0 ≤ 𝑝𝑖 ≤ 1

• Multiple solutions

DDDAS 201725

New Formulation

• Assume Boolean random variables

are independent

• Express each disjunction clause as:

P(AvB) = P(A) + P(B) – P(A^B)

= P(A) + P(B) – P(A)P(B)

• Develop system of nonlinear

equations and solve for atom

probabilities; use these to solve query

DDDAS 201726

Example: Modus Ponens

DDDAS 201727

Some Observations

• Given n-Modus Ponens:

𝐴1𝐴1 → 𝐴2…

𝐴𝑛−1 → 𝐴𝑛

then standard approach needs 2𝑛

models, we solve it in linear time

DDDAS 201728

Some Observations

• Consider Russell & Norvig Wumpus

World:

DDDAS 201729

With logical variables:

• breeze: e.g., B11

• Gold: e.g., G23

• Pit: e.g., P22

• Stench: e.g., S44

• Wumpus: e.g., W34

Rules like: ¬P21 v B11

16*5 = 80 variables

Given the rules, there are:

583 sentences

Some Observations (cont’d)

DDDAS 201730

Breezes Gold Pits Stench Wumpus

Some Observations (cont’d)

DDDAS 201731

URBAN: Uncertainty Reduction-

Based Agent Network

• A mult-agent system specifically designed for geospatial-temporal analysis

across massive distributed datasets.

• Leverages the GeoWave project developed at the National Geospatial-

Intelligence Agency (NGA) (http://locationtech.github.io/geowave/) and the

open source frameworks Apache Hadoop (for distributed processing) and

Accumulo (for key/value database storage).

• Conceptually, a layer atop GeoWave that provides probabilistic logical

reasoning over space and time.

• Dissemination of knowledge in the form of probabilistic sentences and

maps published to GeoServer (http://geoserver.org/)

• Addresses tasking, processing, exploitation, and dissemination of data

(TPED) with an agile sensor network and the unifying concept of

uncertainty reduction.

DDDAS 201732

http://locationtech.github.io/geowave/

http://geoserver.org/

URBAN Implementation

DDDAS 201733

:Mission PlannerRRT* Planner

:UAV Manager

:Weather Monitor

FExample Instantiations of the BRECCIA Agent

:User

DB

GeoServer

Accumulo

Hadoop

• The BRECCIA Agent represents the

core abstraction for all agents in the

system.

• Agents are distributed across

specialized machines such as UAVs,

mobile laptops, or high performance

computers.

• The inherited components of each

BRECCIA agent enable an overall

system that is dynamic and data-

driven.

BRECCIA Agent

P.L. Logic Module

GeoWave Connector

Specialized Functions

Uncertainty Reduct. Goal

BDI Engine

URBAN

Implementation

34

• The Belief-Desire-Intention (BDI) engine serves a dual purpose • As a software architecture it facilitates the discussion and design of agents

• As a software cognitive model it enables goal-driven behavior (and in our particular implementation data-driven).

• Jason (http://Jason.sourceforge.net/wp/) provides the language interpreter and BDI

engine to BRECCIA agents.

BRECCIA Agent

P.L. Logic Module

GeoWave Connector



BDI Engine

The Jason Reasoning Cycle. From Programming Multi-Agent Systems in AgentSpeak

Using Jason (pg. 68), by Rafael H. Bordini et al., 2007, England: John Wiley and Sons Ltd.

http://jason.sourceforge.net/wp/


35

BRECCIA Agent

P.L. Logic Module

GeoWave Connector



BDI Engine


36

• How does Jason enable data-driven behavior?• Plans are executed due to events which may be achievement requests or a change in belief.

• Example: Consider the case where a UAV is executing a path and periodically querying the

geospatial database for path obstruction. To cause the agent to re-plan in the event of an

obstruction, the code is as follows:

+path_obstructed(PathName) -> !replan(PathName)

• The language defined by Jason is inherently data-driven.

BRECCIA Agent

P.L. Logic Module

GeoWave Connector



BDI Engine

React to the belief that path is obstructed… …by replanning


37

• How do Uncertainty Reduction and Probabilistic Logic achieve an agile sensor

network and TPED?• Probabilistic logic and quantified uncertainty provide a means for applying well-known planning

algorithms to an abstract problem space.

• Certainty as a reward for plans:

• Rewards are higher for truth and certainty of positive sentences, or falsity and certainty for

negative sentences.

• Rewards are lowest for uncertain sentences, i.e. p=0.5.

• Non-biased sentences (i.e. inference rules) have a symmetrical reward function.

• The uncertainty reduction goal is executed by a plan to maximize the certainty reward

BRECCIA Agent

GeoWave Connector



P.L. Logic Module

BDI Engine


38

• Uncertainty Reduction Example• Consider the case where an analyst is cooperating with a UAV to gather information about a

location. Associated with a path plan for the UAV are the following sentences and inference for

mission success:

Plan A: a:path_obstructed[p=0.1], b:battery_ok[p=0.9], c:target_recorded[p=0.7], b^c -> mission_success [p=0.84]

• During the course of the mission, a second analyst, cooperating on an unrelated mission, reports seeing smoke at

a location that crosses Plan A’s path. Belief “a” updates, generates an event and causes the UAV agent to replan

– in this case two plans are generated, one to continue on with the mission and another to turn back.

Plan B: a:path_obstructed[p=0.1], b:battery_ok[p=0.8], c:target_recorded[p=0.7], b^c -> mission_success [p=0.82]

Plan C: a:path_obstructed[p=0.05], b:battery_ok[p=0.95], c:target_recorded[p=0.2], b^c -> mission_success [p=0.47]

BRECCIA Agent

GeoWave Connector



P.L. Logic Module

BDI Engine

Target

UAV

Plan A

Plan B

Plan C

Legend


39

• Uncertainty Reduction ExamplePlan A: a:path_obstructed[p=0.6], b:battery_ok[p=0.9], c:target_recorded[p=0.7], b^c -> mission_success [p=0.84]

Reward: R(a,0.6) + R(b,0.9) + R(c,0.7) + R(0.84) = 0.3

Plan B: a:path_obstructed[p=0.1], b:battery_ok[p=0.8], c:target_recorded[p=0.7], b^c -> mission_success [p=0.82]

Reward: R(a,0.1) + R(b,0.8) + R(c,0.7) + R(0.82) = 0.45

Plan C: a:path_obstructed[p=0.05], b:battery_ok[p=0.95], c:target_recorded[p=0.2], b^c -> mission_success [p=0.47]

Reward: R(a,0.05) + R(b,0.95) + R(c,0.2) + R(0.47) = 0.42

• Although the probability of mission success reduced by a small amount due to the

uncertainty in battery usage, the overall uncertainty for this set of sentences has

decreased by choosing Plan B.

BRECCIA Agent

GeoWave Connector



P.L. Logic Module

BDI Engine

Target

UAV

Plan A

Plan B

Plan C

Legend


40

• Uncertainty Reduction Example• Demonstrates how uncertainty reduction and probabilistic logic

facilitate planning over heterogeneous data – paths, events, human

observations.

• Provides justification for decision makers – the agent may request

guidance from the user if the plan alternatives are close to one

another

BRECCIA Agent

GeoWave Connector



P.L. Logic Module

BDI Engine


41

• GeoWave Connector• GeoWave enables agents to simultaneously access a distributed geospatial-temporal

database.

• Agents publish geospatial knowledge, written to the database, via GeoServer. This

enables remote sharing of this type of knowledge.

• In Jason, internal actions coded into the GeoWave connector provide direct access to

the databases.

• Example from weather agent:

+!share_storm_info(Location, Agent) ->

geowaveConnector::get_wms_url(Location, WmsUrl) ;

.send(Agent, tell, storm_info(WmsUrl).

• Data-driven response from UAV agent:

+storm_info(WmsUrl) -> !check_path_obstruction(WmsUrl)

BRECCIA Agent

P.L. Logic Module

GeoWave Connector



BDI Engine


42

• Distributed approach enables web-based user access:

BRECCIA Agent

P.L. Logic Module

GeoWave Connector



BDI Engine

Prototype BRECCIA Client interface and Chat Window Featuring Map of Salt Lake City from

Local GeoServer Instance


43

• Specialized Functions

• Current implementations of specialized functions include

• Connecting to MATLAB instances (Agents who know how to use MATLAB)

• RRT* path planner (Agents who know how to plan over space with vehicle

constraints)

• Wind Simulator (Agent that runs a wind vortex simulator)

• Ongoing work of specialized functions

• GDELT database query (Agents that can query the massive GDELT global event

database (http://www.gdeltproject.org/)

• OpenWeather API Agent (Agents that can query distributed weather information)

• UAV simulator (Agents that can run real-time UAV simulators)

• UAV controller (Agents that can control quadcopters in real-time)

BRECCIA Agent

P.L. Logic Module



BDI Engine

GeoWave Connector

http://www.gdeltproject.org/

Conclusions

• Developed effective and efficient

probabilistic logic method

• Developed core of BRECCIA system

• URBAN: allows communicating,

autonomous agents

• Cloud computing

DDDAS 201744

Future Work

• Probabilistic Logic

• Use conditional probabilities

• P(A&B) = P(A|B)P(B)

• approximate P(𝐴1& 𝐴2 & … & 𝐴𝑛) by upper

bound:

• min(P(𝐴𝑖 & 𝐴𝑗)), 𝑖, 𝑗 = 1: 𝑛

• Extend to First Order Logic

• Exploit argumentation to reduce

analysis cost

DDDAS 201745

Future Work

• BRECCIA middleware

• Extend breadth of agents

• Extend depth of analysis

• Develop realistic ISR scenario

• Collaborate with Air Force ISR wing

• Measure performance

• Accuracy of uncertainty measures

• Time & Space Computational aspects

DDDAS 201746

Future Work

• Simulation

• Use full 3D QUIC-URB modeling system

• Exploit Gaussian process models of 3D

features

• Terrain

• Temperature

• Wind

• Obscurants

DDDAS 201747

Future Work

• Experiments

• Develop real-world ISR missions

• Small-scale experiments in lab

• Full-scale experiments (in SLC)

• Measures of performance (MOP)

& Effectiveness(MOE)

• Success (MOP)

• Time (MOP)

• Information adequacy (MOE)

DDDAS 201748

UTAH UAV Fleet

DDDAS 201749

Questions?

DDDAS 201750

BRECCIA: A Multi-Agent Data Fusion and Decision …...BRECCIA Agent P.L. Logic Module GeoWave...

Documents

Transcript of BRECCIA: A Multi-Agent Data Fusion and Decision …...BRECCIA Agent P.L. Logic Module GeoWave...