1 of 12

Characterizing Novelty �in the DOD Domain

27^th ICCRTS

25-27 October 2022 - Quebec City (Canada)

Presented by Theresa Chadwick¹ (theresa.m.chadwick.civ@us.navy.mil)

Co-contributors: Nicholas Soultanian¹, Douglas Lange¹, and Hector Ortiz-Pena²

1. Naval Information Warfare Center Pacific, San Diego, California

2. CUBRC, Buffalo, New York

10/22/22

Distribution Statement A. Approved for public release: distribution unlimited.

2 of 12

Outline

Introduction to Novelty
SAIL-ON Overview
Effects of Novelty on an AI Agent
Theory of Novelty
Defining Novelty within the DOD
Mathematically Modeling Novelty
Generation of Novel Battles

10/22/22

Distribution Statement A. Approved for public release: distribution unlimited.

3 of 12

Introduction

What is novelty?

“We use the term “novelty” here to refer to situations that violate implicit or explicit assumptions about agents, the environment, or their interactions.” SAIL-ON BAA

What does novelty mean for the warfighter?

Novelty might be best described as something unexpected or not considered given previous training and intelligence.
Examples: NATO forces conducted an aerial bombing against Serbia during the Kosovo War, called Operation Allied Force.
Example: Black Swan Events such as September 11, 2001

Goal: To characterize novelty within the DOD Domain and use those characterizations to test AI agents against in a simulated environment.

10/22/22

Distribution Statement A. Approved for public release: distribution unlimited.

4 of 12

SAIL-ON Overview

NIWC Pacific

Distribution Statement A. Approved for public release: distribution unlimited.

5 of 12

Effects of Novelty on AI Agents

Currently, AI agents struggle with handling inputs it has not seen before. One of the goals of SAIL-ON is to create robust AI agents with the capability to detect, characterize, and accommodate novelties in real time.

10/22/22

Enemy SAM Site

Enemy Ammo Storage Site (Target)

Blue Fighter Jet (AI agent)

Pre-Novelty Missile Range

Post-Novelty Missile Range

Distribution Statement A. Approved for public release: distribution unlimited.

To further illustrate this point, we will refer to a simulated tactical scenario called Scenario Zero in which a blue force fighter jet (controlled by an AI agent), which helps make its decision based on various sensor data, is tasked to destroy a red enemy’s ammo storage site while evading two nearby red enemy force’s surface-to-air missile (SAM) launcher. In a pre-novelty scenario, the SAM’s missile range is held constant, and the AI agent can successfully execute its task without having the fighter jet shot down. However, when considering the rapid advancements in technology, it is an obvious oversight to assume that the SAM’s missile range will always be held constant. In the post-novelty, open world scenario, the SAM’s range is increased beyond what the blue fighter jet was previously accustomed to, causing the agent to be shot down before executing its task.
Rather than replanning its course to avoid being shot down, the AI agent in Scenario Zero does not even recognize that a novel situation has occurred, and it continues to send more assets to carry out its task only for them to be destroyed. In an ideal scenario where AI agents are created to be aware of the possibility of novelty, a more robust agent could instead learn to detect, characterize, and accommodate real world novelties.

6 of 12

Capstone Scenario

NIWC Pacific

Yellow Dash line indicates actual SAM range.

(i.e. actual missile range)

Blue Force Risk Assessment Heat Map

BLUE Perception of SAM Weapon Exclusion Zone

(i.e. perceived missile range)

Assumption about RED capabilities is encoded into BLUE Risk Assessment

High Risk




Low Risk

Distribution Statement A. Approved for public release: distribution unlimited.

7 of 12

Theory of Novelty

DARPA’s Novelty Working Group came up with a Novelty Hierarchy to classify novelties depending on the complexity of the domain it affects.
We use this hierarchy to help define novelty within the DOD.
Some important notes:

Each level also contains new classes, attributes, or representations.
Avoid “nuisance” novelty.
Difficulty of detection, characterization, and accommodation does not necessarily increase with each level.

10/22/22

Single Entities	Phase 1	1	Objects: New classes, attributes, or representations of non-volitional entities.
	Phase 2	2	Agents: New classes, attributes, or representations of volitional entities.
		3	Actions: New classes, attributes, or representations of external agent behavior.
Multiple Entities		4	Relations: New classes, attributes, or representations of static properties of the relationships between multiple entities.
Multiple Entities		5	Interactions: New classes, attributes, or representations of dynamic properties of behaviors impacting multiple entities.
Complex Phenomena	Phase 3	6	Rules: New classes, attributes, or representations of global constraints that impact all entities.
		7	Goals: New classes, attributes, or representations of external agent objectives.
		8	Events: New classes, attributes, or representations of series of state changes that are not the direct result of volitional action by an external agent or the SAIL-ON agent.

Table 1: Open World Novelty Hierarchy [6]

Distribution Statement A. Approved for public release: distribution unlimited.

8 of 12

Defining Novelty within the DOD

10/22/22

Novelty Level	Examples
Object	Addition of nearby no-fire entities, such as civilian towns or embassies. Arduous or view-limiting terrain.
Agent	Advancing the capabilities of one or more SAMs, such as increased missile range.
Action	One or more SAMs become mobile or a decoy.
Relation	Changing the location of the SAM to a more difficult terrain or closer to no-fire entities. One or more sensors break and transmit no data to the fighter jet.
Interaction	Survivability of SAM or ammo storage site increases. One or more sensors breaks or becomes compromised to convey false data.
Rules	No-fire zones for all entities. Every entity has a time delay when executing an order to fire.
Goals	One or more of the SAM sites are no longer tasked with the mission defending the ammo storage site. Instead they are tasked with the mission of destroying the fighter jet.
Environment	Introduction of a red enemy fighter jet in the middle of the scenario.

Distribution Statement A. Approved for public release: distribution unlimited.

9 of 12

Mathematically Modeling Novelty

10/22/22

SAM Site

Pre-Novelty

Easy Distribution

Medium Distribution

Hard Distribution

Distribution Statement A. Approved for public release: distribution unlimited.

10 of 12

Generation of Novel Battles

CUBRC has developed the Simulation Orchestration for Learning and Validation Environment (SOLVE)

Can leverage video games and military simulators to rapidly test and evaluate mission planning algorithms.
Currently integrates with the Advanced Framework for Simulation, Integration, and Modeling (AFSIM) and One Semi-Automated Forces (OneSAF).

We will use SOLVE to create Campaigns of Battles to test novelties against different AI agents.
We will compare AI agents against metrics based on mission performance and detection, characterization, and accommodation values.

10/22/22

Distribution Statement A. Approved for public release: distribution unlimited.

11 of 12

Acknowledgements

This research was sponsored by DARPA.
The following teams and individuals are part of the NWG and contributed to the open-world novelty hierarchy presented in Table 1:

10/22/22

Institute for Defense Analysis led by Josh Alspector and Pat Langley
Washington State University led by Lawrence Holder
Australian National University led by Jochen Renz
University of Southern California led by Mayank Kejriwal
University of Maryland led by Abhniav Shrivastava
University of Texas at Dallas led by Eric Kildebeck
University of Massachusetts at Amherst led by David Jensen
Tufts University led by Matthias Scheutz
Rutgers led by Patrick Shafto
Georgia Tech led by Mark Riedl
PAR Government led by Eric Robertson
SRI International led by Giedrius Burachas
Charles River Analytics led by Bryan Loyall
Xerox PARC led by Shiwali Mohan
Smart Information Flow Technologies led by David Musliner
Raytheon BBN Technologies led by Bill Ferguson
Kitware led by Anthony Hoogs
Tom Dietterich
Marshall Brinn
Jivko Sinapov

Distribution Statement A. Approved for public release: distribution unlimited.

12 of 12

References

[1] Alspector, J. “Representation Edit Distance as a Measure of Novelty,” in Proceedings of AAAI Spring Symposium on Designing Artificial Intelligence for Open Worlds, 2022, https://usc-isi-i2.github.io/AAAI2022SS/papers/SSS-22_paper_18.pdf
[2] Boult, T. E., et al. “A Unifying Framework for Formal Theories of Novelty: Framework, Examples, and Discussion.” ArXiv abs/2012.04226 (2020). https://arxiv.org/pdf/2012.04226.pdf
[3] Boult, T., et al. (2021). “Towards a Unifying Framework for Formal Theories of Novelty,” in Proceedings of the AAAI Conference on Artificial Intelligence, 35(17), 15047-15052. Retrieved from https://ojs.aaai.org/index.php/AAAI/article/view/17766
[4] Burns, W. R. and Miller, D. “Improving DOD Adaptability and Capability to Survive Black Swan Events.” National Defense University Press. ndupress.ndu.edu. Q1 2014. https://ndupress.ndu.edu/Media/News/Article/577479/improving-dod-adaptability-and-capability-to-survive-black-swan-events/#:~:text=1%20A%20black%20swan%20event,is%20not%20an%20unpredictable%20event
[5] Chao, J., et al. “Evaluating Reinforcement Learning Algorithms For Evolving Military Games,” in Proceedings of AAAI Symposium on The 2nd Workshop on Deep Models and Artificial Intelligence for Defense Applications: Potentials, Theories, Practices, Tools, and Risks, 2020, http://ceur-ws.org/Vol-2819/session3paper2.pdf
[6] Doctor, K., et al. “Toward Defining a Domain Complexity Measure Across Domains,” in Proceedings of AAAI Spring Symposium on Designing Artificial Intelligence for Open Worlds, 2022, https://usc-isi-i2.github.io/AAAI2022SS/papers/SSS-22_paper_79.pdf
[7] Doctrine for the Armed Forces of the United States, Joint Publication 1, USA., 2013, pp. 12. https://www.jcs.mil/Portals/36/Documents/Doctrine/pubs/jp1_ch1.pdf. (March 3, 2022)
[8] Lambeth, B. S., “Operation Allied Force: Lessons for the Future.” Santa Monica, CA: RAND Corporation, 2001. https://www.rand.org/pubs/research_briefs/RB75.html
[9] Muhammad, F., et al. “A novelty-centric agent architecture for changing worlds,” in Proceedings of 20th International Conference on Autonomous Agents and Multiagent Systems, 2021. https://hrilab.tufts.edu/publications/muhammadetal21aamas.pdf
[10] Pinot, V., et al. “Measuring the Performance of Open-World AI Systems” in Proceedings of AAAI Spring Symposium on Designing Artificial Intelligence for Open Worlds, 2022, https://usc-isi-i2.github.io/AAAI2022SS/papers/SSS-22_paper_73.pdf
[11] Senator, T. 2019. Science of Artificial Intelligence and Learning for Open-world Novelty (SAIL-ON). https://www.darpa.mil/program/science-of-artificialintelligence-and-learning-for-open-world-novelty.
[12] Silver, D., et al. “Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm.” ArXiv abs/1712.01815 (2017). https://arxiv.org/pdf/1712.01815.pdf.
[13] Xue, C., Pinto, V., et al. “Science Birds Novelty: an Open-world Learning Test-bed for Physics Domains,” in Proceedings of AAAI Spring Symposium on Designing Artificial Intelligence for Open Worlds, 2022, https://usc-isi-i2.github.io/AAAI2022SS/papers/SSS-22_paper_75.

10/22/22

Distribution Statement A. Approved for public release: distribution unlimited.