SLIDE 1
Page 2
COMMERCIAL IN CONFIDENCE
Agenda
- Background
– Genesis – Concept
- Approach
– Architecture – Testing – Red AIs
- Results
– ESRA – DR SO
- Prediction accuracy
- Identifying unknowns
- Next steps
Red Mirror: Counter AI AI UDT 2020, Rotterdam Ahoy, NL 27 th May - - PowerPoint PPT Presentation
Red Mirror: Counter AI AI UDT 2020, Rotterdam Ahoy, NL 27 th May - - PowerPoint PPT Presentation
Red Mirror: Counter AI AI UDT 2020, Rotterdam Ahoy, NL 27 th May 2020 COMMERCIAL IN CONFIDENCE Agenda Background Genesis Concept Approach Architecture Testing Red AIs Results ESRA DR SO Prediction
SLIDE 2
SLIDE 3
Page 3
COMMERCIAL IN CONFIDENCE
Background | Genesis
- If you can explain our AI, can you predict their AI?
SLIDE 4
Page 4
COMMERCIAL IN CONFIDENCE
Background | Concept
- Aims
– Use a combination of Artificial Intelligence (AI) and Machine Learning (ML) techniques to automatically generate a ‘mirror’ of Red’s AI i.e. the AI controlling enemy vessels
- Outputs
– Predict Red’s next Courses Of Action (COA) – Identify what Red knows that Blue does not – Identify what Blue knows that Red does not – Suggest what Blue should do to build a better Red Mirror (incl ‘he knows we know he knows…’ effects) – Explain the Red-AI behaviour so that the Intelligent Core or human decision-maker can check whether the Red-AI may be trying to double-bluff
- Potential use-cases
– Turning Red’s assets into a form of ‘sensor input’ for Blue – Getting ahead of Red i.e. SUPA loop…
SLIDE 5
Page 5
COMMERCIAL IN CONFIDENCE
Background | Concept
SLIDE 6
Page 6
COMMERCIAL IN CONFIDENCE
Approach | Architecture
SLIDE 7
Page 7
COMMERCIAL IN CONFIDENCE
Approach | Testing
- Measures of performance
– Accuracy of predicting Red’s next COA
- Over time/observations
- By specific action
- By degree
– Uncertainty of prediction
- By probability
- By degree
– Accuracy of identifying what Red knows that Blue does not – Accuracy of identifying what Red knows that Blue does not
- Test cases
– Different ‘ Red AIs’
- ESRA
- DR SO
– Knowledge
- Full
- Full for Red not for Blue
- Full for Blue not for Red
SLIDE 8
Page 8
COMMERCIAL IN CONFIDENCE
Approach | Red AIs
- ESRA
– 1 red ship avoiding and destroying 8 blue missiles)
No_Of_ Tracks Missile_ Number Threat_L evel_Of_ Missile Bearing _Of_Mis sile Type_Of _Missile Speed_ Of_Miss ile Missile_ Engaged Course_Of_Action 4 7 0.5 300 KH-35U 434 7 HK1 track 0 4 8 0.4 30 KH-35U 434 8 SK0 track 0 4 5 0.7 30 KH-31A 600 5 HK1 track 0 4 8 0.3 120 KH-35U 434 N/A Wait
SLIDE 9
Page 9
COMMERCIAL IN CONFIDENCE
Approach | Red AIs
- DR SO
– 4 Red agents chasing 1 Blue agent, around obstacles
SLIDE 10
Page 10
COMMERCIAL IN CONFIDENCE
Results | ESRA – prediction accuracy vs categorical outputs
SLIDE 11
Page 11
COMMERCIAL IN CONFIDENCE
Results | DR SO – prediction accuracy vs continuous outputs
SLIDE 12
Page 12
COMMERCIAL IN CONFIDENCE
Results | DR SO – confidence in prediction accuracy
SLIDE 13
Page 13
COMMERCIAL IN CONFIDENCE
Results | DR SO – what Blue does not know
SLIDE 14
Page 14
COMMERCIAL IN CONFIDENCE
Next steps
- Short-term
– Demonstrate the operational impact of the demonstrated prediction accuracy, and accuracy of identifying the Blue or Red unknown e.g.
- How much extra time does the Blue agent get before the DR SO Red agents successfully surround or
swarm it?
- How much does the Blue probability of success increase against a Red AD platform using ESRA, in
terms of kill probability or rate of reduction of Red stocks? – Test a ‘reverse-decision tree’ or ‘reverse random forest’ method for identifying unknown information to improve the balance between accuracy of identification and FPR – Testing methods for extracting, or back calculating, the reward structure of the Red AI, which is implicit in the Red Mirror
- Medium-term
– Wider project to explore some of the issues more deeply e.g.
- What types of Red AI we would wish to apply the Red Mirror concept to and what benefits accrue?
- How is the difficulty of predicting AI COAs driven by the COAs and the scenarios it is used in?
- Why might it be correct to not use all the past observations to predict Red, and how might the
threshold after which past data is useful be found?
- Why might partial knowledge improve short-term prediction accuracy but worsen longer-term
prediction accuracy?
- How might deception be exploited in counter-AI activity?