AI on the Battlefield: an Experimental Exploration Alexander Kott BBN Technologies Robert Rasch US Army Battle Command Battle Lab Views expressed in this paper are those of the authors and do not necessarily reflect those of the U. S. Army or any agency of the U.S. government. Kenneth Forbus Northwestern University
Outline Motivation for the experiment The experimental rig Experimental procedure Findings A surprising challenge uncovered
The Role of BCBL-L Exploration of new techniques and tool for Army C2 – a key focus of BCBL-L Apparent emergence and maturing of multiple technologies for MDMP What is the right way to apply such technologies? Value? Drawbacks? BCBL-L proposed and executed the Concept Experimentation Program (CEP) - Integrated Course of Action Critiquing and Elaboration System (ICCES)
Room for Controversy Some call for “…fast new planning processes… between man and machine… decision aids…” Extensive training and specialization requirements? Detract from intuitive, adaptive, art-like aspects of military command? Undue dependence on vulnerable technology? Make the plans and actions more predictable to the enemy? The experiment was designed to address such concerns
The Experimental Rig COA Creator, by the Qualitative Reasoning Group at Northwestern University - allows a user to sketch a COA The COA statement tool, by Alphatech, allows the user to enter the COA statement Fusion engine, by Teknowledge, fuses the COA sketch and statement CADET, by Carnegie Group & BBN – elaborates the fused sketch- and-statement into a detailed plan and estimates Input: Mission and Intelligence Analysis CADET Tool Fusion Tool COA Statement Tool COA Creator Tool Output: Detailed Synchron. Matrix
The COA Entry Bottleneck The key bottleneck in MDMP digitization: Time / effort / distraction Training requirements Downstream representation language Our approach – COA Creator, based on nuSketch Sketching = interactive drawing plus linguistic I/O Rich conceptual understanding of the domain Speech often not preferred in mix of modalities Include “speechless” multimodal interface (buttons plus gestures) Expressible in the underlying knowledge representation
Terrain features and characterization
Units and control lines
Objective and engagement areas
Friendly tasks are defined
The Experimental Procedure Comparison with the conventional process Exploratory vs. statistical rigor Training Interviews, Products Review Team 1, Case 2Team 2, Case 2 Team 2, Case 1Team1, Case1 Conventional Manual Process ICCES- Based Process
Key Findings Low training requirements Largely due to “naturalness” of sketching Simple, frugal CONOPS No impact on creative aspects of the process Largely driven by human-generated sketch-and-statement Opportunity to explore more options Dramatic time savings (3-5 times faster) Mainly in downstream processing (e.g., planning) Comparable quality of products Few edits of ICCES-built products Comparable quantitative measures (e.g., friendly losses)
Parallel Experiments – Quality of Plans Rigorous experimental comparison: computer-assisted vs. conventional Multiple cases, subject, judges Conclusions: indistinguishable quality of products, dramatically faster Products of 5 past exercises Grade by 9 “Blind” Judges Generate w/ CADET Give “computer look” inputs outputs
Surprise: Plan Presentation is a Key Concern Conventional output presentation paradigms, i.e., sync. matrix is ineffective Larger number of elements Inadequate spatial aspect Difficult to detect errors Alternatives: Animation? Cartoon sketches?
Conclusions For Army professionals: Technologies like ICCES have near-term deployment potential No impact on creativity, predictability Dramatic acceleration, comparable quality Challenges in inspecting, comprehending the new MDMP products For AI R&D community: Dominant role of HMI challenges calls for new mechanisms Value of natural sketch- based interfaces Simple, straightforward, all- in-one CONOPS for users No substitute for comparative experiments, from both practical and research perspectives
BACKUP SLIDES