Download presentation
Presentation is loading. Please wait.
1
Cranfield Universityb (UK)
UNCLASSIFIED Challenges and Opportunities in Applying Bayesian Network Modelling to Strategic Force Design Thang Caoa, Andrew Couttsa, Ben Pietscha, Dermot Blumsona and Adam Zagoreckib Defence Science & Technology Groupa (Australia) Cranfield Universityb (UK)
2
Outline Problem Statement and Motivation
UNCLASSIFIED Outline Problem Statement and Motivation Modelling approach for evaluating force design options Qualitative, quantitative and Bayesian Networks DeMorgan qualitative Bayesian Network model Causal Strength Logic (CAST) Challenges Future work and discussion
3
Strategic force design problem statement and motivation
UNCLASSIFIED Strategic force design problem statement and motivation Scale – whole of force to tactical unit design Temporal – now to the next 30 years Uncertainty – environment, technology, force effectiveness ect Complexity – cause and effect relationships Evaluation – compare design options Methodology – qualitative, quantitative methodology or both Motivation: Developing a modelling framework to address above problems
4
Example: US Army Design Methodology
UNCLASSIFIED Example: US Army Design Methodology “methodology for applying critical and creative thinking to understand, visualise, and describe unfamiliar problems and approaches to solving them”, Department of the Army. (2008). TRADOC Pamphlet , Commander’s Appreciation and Campaign Design, Fort Monroe, VA: Headquarters, United States Army “systemic rather than reductionist, and qualitative rather than quantitative” The Art of Design, Military Review; Ryan, A. J., Schifferle, P., Stewart, M., Butler-Smith, A., Schmidt, M., Rochelle, J. & Webb, G. (2010). Figure 5. Original visualisation of the U.S. Army Design Methodology.
5
Systemic Design Influence Diagram for force design
Unclassified Systemic Design Influence Diagram for force design
6
Force design evaluation model hierarchy
UNCLASSIFIED Force design evaluation model hierarchy
7
Qualitative vs quantitative evaluation of options
UNCLASSIFIED Qualitative vs quantitative evaluation of options For qualitative evaluation of options: Accounts for strategic factors influencing selection Enables complex details within an overall expert assessment Less preparation or agreement between stakeholders on methods and data Against qualitative evaluation of options: Subjective – Difficult for aggregation and repeatability Slow – Difficult to conduct rapid ‘What if’ analysis Opaque – Assessments may not provide full detail justification for result Synthesis qualitative and quantitative Support qualitative analysis with rapid structured/quantitative options comparison Provide transparent explanation for assessments within the decision system Propose Bayesian Network (BN) as a methodology to achieve this
8
Building a BN Model for evaluating force options
UNCLASSIFIED Building a BN Model for evaluating force options BNs are a practical modeling tool – but efficient model building techniques are needed Possible approaches to building models: Use human expert(s) Directly from data Combination of expert’s knowledge and data Challenges: Capture complex relationship graphically Large number of probabilities required to specify conditional probability tables – growth exponentially – difficult for elicitation Adapting elicitation for elites and non-technical experts
9
Reduce size and elicitation of CPTs for qualitative problem
UNCLASSIFIED Reduce size and elicitation of CPTs for qualitative problem There are at least three different approaches to the problem: Independence of causal influences models (ICI) assume some model of interactions between causes and the effect Context specific independence (CSI) exploit and encode efficiently independences (symmetries) in CPTs Bayesian Network Interpolation Model (Cain’s elicitation calculator) Exploit interpolation and parent interaction property to reduce the number of elicitations There is no formal classification in the literature and there are examples that do not strictly fit any of the three approaches
10
Qualitative BNM from DeMorgan’s Law
UNCLASSIFIED Qualitative BNM from DeMorgan’s Law BN DeMorgan Model Assume binary states and ICI Modelling four types of cause-effect relationship: Cause, Barrier, Requirements and Inhibitor (16 Boolean functions) Derive conditional probabilities from linearly growth elicited CPT Model both positive and negative influence Pros: Easy to interact with experts, good probabilistic semantics, model unobservable cause-effect, linearly growth elicited CPT Cons: Binary states, require ICI assumption Example: Red to green colour indicates low to high probability
11
Qualitative BN from Causal Strength (CAST) Logic
UNCLASSIFIED Qualitative BN from Causal Strength (CAST) Logic CAST Algorithm Aggregate positive causal strengths Aggregate negative causal strengths Combine positive and negative causal strengths, and Derive conditional probabilities Pros: Easy to interact with experts, proven successful with many US military applications Cons: Binary states, lack of probabilistic semantics, require ICI assumption Example: Red to green colour indicates low to high probability
12
UNCLASSIFIED Prototype: Land force design qualitative Bayesian network model (DeMorgan) Red to green colour indicates low to high probability
13
Proposed force design evaluation framework
UNCLASSIFIED Proposed force design evaluation framework Qualitative Bayesian Network Model (QBNM) for evaluating high level force design Apply QBNM (DeMorgan or CAST)for rapidly evaluating strategic factors Support qualitative analysis with rapid structured/quantitative options comparison Extending QBNM to full CPT BNM with multiple states Extending BNM to multiple states Applying interpolation algorithm for CPT gap-filling and allowing interaction between parent nodes . Provide transparent decision model through Bayesian Decision Theory
14
Conclusion & Future Work
UNCLASSIFIED Conclusion & Future Work The proposed framework addressed: Uncertainty, complexity, evaluation and methodology Scale – on high level force design only Temporal – need to extend to Dynamic Bayesian Network Model (DBNM) Future work Methodology and model validation (applied to force design problem only) Extend DeMorgan and CAST models to multiple states Investigate the applicability of CSI model Develop BN meta-model from simulation for modelling tactical unit force design Extend to DBNM for evaluating temporal force design Collaboration
15
UNCLASSIFIED Questions ?
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.