s ystems analysis laboratory helsinki university of technology games and bayesian networks in air...
TRANSCRIPT
S ystemsAnalysis LaboratoryHelsinki University of Technology
Games and Bayesian Networks Games and Bayesian Networks in Air Combat Simulation Analysisin Air Combat Simulation Analysis
M.Sc. Jirka Poropudas and Dr.Tech. Kai Virtanen
Systems Analysis LaboratoryHelsinki University of Technology
S ystemsAnalysis LaboratoryHelsinki University of Technology
OutlineOutline
• Air combat (AC) simulation
• Games in validation and optimization– Estimation of games from simulation data
– Analysis of estimated games
• Dynamic Bayesian networks (DBNs)– Estimation of DBNs from simulation data
– Analysis of estimated DBNs
• Conclusions
S ystemsAnalysis LaboratoryHelsinki University of Technology
Air Combat SimulationAir Combat Simulation
• Commonly used models based on discrete event simulation
• Most cost-efficient and flexible method
Objectives for AC simulation studies: • Acquire information on systems performance• Compare tactics and hardware configurations• Increase understanding of AC and its progress
S ystemsAnalysis LaboratoryHelsinki University of Technology
Discrete Event Simulation ModelDiscrete Event Simulation Model
Simulation input• Aircraft and
hardware configurations
• Tactics
• Decision making parameters
Simulation output• Number of kills and
losses
• Aircraft trajectories
• AC events
• etc.Decision making logic
Aircraft, weapons and hardware models
Stochastic elements
Validation of the model?Optimization of output?Evolution of simulation?
S ystemsAnalysis LaboratoryHelsinki University of Technology
Existing Approaches to Simulation AnalysisExisting Approaches to Simulation Analysis
• Simulation metamodels– Mappings from simulation input to output
- Response surface methods, regression models, neural networks
• Validation methods– Real data, expert knowledge, statistical methods, sensitivity
analysis
• Simulation-optimization methods– Ranking and selection, stochastic gradient approximation,
metaheuristics, sample path optimization
S ystemsAnalysis LaboratoryHelsinki University of Technology
Limitations of Existing ApproachesLimitations of Existing Approaches
• Existing approaches are one-sided– Action of the adversary is not taken into account– Two-sided setting studied with games
• Existing approaches are static– AC is turned into a static event– Time evolution studied with dynamic Bayesian networks
S ystemsAnalysis LaboratoryHelsinki University of Technology
Games from Simulation DataGames from Simulation Data• Definition of scenario
– Aircraft, weapons, sensory and other systems– Initial geometry– Objectives = Measures of effectiveness (MOEs)– Available tactics and systems = Tactical alternatives
• Simulation of the scenario– Input: tactical alternatives– Output: MOE estimates
• Games estimated from the simulation data• Games used for validation and/or
optimization
S ystemsAnalysis LaboratoryHelsinki University of Technology
Estimation of GamesEstimation of Games
RED
0.0040.036-0.833x3
0.0230.013-0.811x2
0.8850.855-0.077 x1
y3y2y1
BL
UE
IIIIIIIx3
IIIIIIIx2
IVIVII x1
y3y2y1
RED, min
BL
UE
, max
MOE estimates Payoff
Discrete tactical alternatives x and y
Analysis of variance
Simulation
Discrete decision variables x and y
Game
S ystemsAnalysis LaboratoryHelsinki University of Technology
Estimation of GamesEstimation of Games
MOE estimates Payoff
Continuous tactical alternatives x and y
Simulation
Continuous decision variables x and y
Game
0
5
10
15
0
5
10
15
0
0.1
0.2
0.3
0.4
0.5
0
5
10
15
0
5
10
15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Blue xRed y
MO
E
estim
ate
Blue xRed y
Pay
off
Regression analysis
Experimental design
S ystemsAnalysis LaboratoryHelsinki University of Technology
Analysis of GamesAnalysis of Games• Validation: Confirming that the simulation model
performs as intended
– Comparison of the scenario and properties of the game
– Symmetry, dependence between decision variables and payoffs, best responses and Nash equilibria
• Optimization: Comparison of effectiveness of tactical alternatives
– Different payoffs, best responses and Nash equilibria, dominance between alternatives, max-min solutions
S ystemsAnalysis LaboratoryHelsinki University of Technology
Example: Missile Support Time GameExample: Missile Support Time Game
yx
x
y
Phase 1: SupportRelay radar information on the adversary to the missile
Phase 2: Extrapolation
Phase 3: Locked
• Symmetric one-on-one scenario• Tactical alternatives: Support times x and y• Objective => MOE: combination of kill probabilities• Simulation using X-Brawler
S ystemsAnalysis LaboratoryHelsinki University of Technology
Game PayoffsGame PayoffsRegression models for kill probabilities:
Probability of Blue kill Probability of Red kill
0
5
10
15
0
5
10
15
0
0.2
0.4
0.6
0.8
0
5
10
15
0
5
10
15
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Blue’s support time x
Blue’s support time x Red’s support time y
Red’s support time y
Payoff: Weighted sum of kill probabilities
• Blue: wB*Blue kill prob. + (1-wB)*Red kill prob.
• Red: wR*Red kill prob. + (1-wR)*Blue kill prob.
• Weights = Measure of aggressiveness
S ystemsAnalysis LaboratoryHelsinki University of Technology
0 5 10 150
5
10
15
Best ResponsesBest ResponsesBest response =
Optimal support time against a given support time of the adversary
Best responses with different weights
Nash equilibria:Intersections of the best responses
WR=0
WR=0.5
WR=0.25
WR=0.75
WB=0.75WB=0.5WB=0.25WB=0
Blue’s support time x
Re
d’s
su
ppo
rt ti
me
y
S ystemsAnalysis LaboratoryHelsinki University of Technology
Analysis of GameAnalysis of Game
• Symmetry– Symmetric kill probabilities and best responses
• Dependency– Increasing support times => Increase of kill probabililties
• Different payoffs– Increasing aggressiveness (higher values of wB and wR)
=> Longer support times• Best responses & Nash equilibria
– Increasing aggressiveness (higher values of wB and wR) => Longer support times
S ystemsAnalysis LaboratoryHelsinki University of Technology
DBNs from Simulation DataDBNs from Simulation Data• Definition of simulation state
– Aircraft, weapons, sensory and other systems
• Simulation of the scenario– Input: tactical alternatives– Output: simulation state at all times
• DBNs estimated from the simulation data– Network structure– Network parameters
• DBNs used to analyze evolution of AC– Probabilities of AC states at time t– What if -analysis
S ystemsAnalysis LaboratoryHelsinki University of Technology
Definition of State of ACDefinition of State of AC
• 1 vs. 1 AC
• Blue and Red
• Bt and Rt = AC state at time t
• State variable values
• “Phases” of simulated pilots
– Part of the decision making model
– Determine behavior and phase transitions for individual pilots
– Answer the question ”What is the pilot doing at time t?” Example of AC phases in X-Brawler
simulation model
S ystemsAnalysis LaboratoryHelsinki University of Technology
Dynamic Bayesian Network for ACDynamic Bayesian Network for AC
• Dynamic Bayesian network– Nodes = variables
– Arcs = dependencies
• Dependence between variables described by– Network structure
– Conditional probability tables
• Time instant t presented by single time slice
• Outcome Ot depends on Bt and Rt
time slice
S ystemsAnalysis LaboratoryHelsinki University of Technology
Dynamic Bayesian NetworkDynamic Bayesian NetworkFitted to Simulation DataFitted to Simulation Data
• Basic structure of DBN is assumed
• Additional arcs added to improve fit
• Probability tables estimated from simulation data
S ystemsAnalysis LaboratoryHelsinki University of Technology
• Continuous probability curves estimated from simulation data
• DBN model re-produces probabilities at discrete times
• DBN gives compact and efficient model for the progress of AC
Evolution of ACEvolution of AC
S ystemsAnalysis LaboratoryHelsinki University of Technology
What If -AnalysisWhat If -Analysis
• Evidence on state of AC fed to DBN
• For example, blue is engaged within visual range combat at time 125 s
– How does this affect the progress of AC?
– Or AC outcome?
• DBN allows fast and efficient updating of probability distributions
– More efficient what-if analysis
• No need for repeated re-screening simulation data
S ystemsAnalysis LaboratoryHelsinki University of Technology
ConclusionsConclusions• New approaches for AC simulation analysis
– Two-sided and dynamic setting
– Simulation data represented in informative and compact form
• Game models used for validation and optimization
• Dynamic Bayesian networks used for analyzing the evolution of AC
• Future research:
– Combination of the approaches => Influence diagram games
S ystemsAnalysis LaboratoryHelsinki University of Technology
References References » Anon. 2002. The X-Brawler air combat simulator management summary. Vienna,
VA, USA: L-3 Communications Analytics Corporation.
» Gibbons, R. 1992. A Primer in Game Theory. Financial Times Prenctice Hall.
» Feuchter, C.A. 2000. Air force analyst’s handbook: on understanding the nature of analysis. Kirtland, NM. USA: Office of Aerospace Studies, Air Force Material Command.
» Jensen, F.V. 2001. Bayesian networks and decision graphs (Information Science and Statistics). Secaucus, NJ, USA: Springer-Verlag New York, Inc.
» Law, A.M. and W.D. Kelton. 2000. Simulation modelling and analysis. New York, NY, USA: McGraw-Hill Higher Education.
» Poropudas, J. and K. Virtanen. 2007. Analyzing Air Combat Simulation Results with Dynamic Bayesian Networks. Proceedings of the 2007 Winter Simulation Conference.
» Poropudas, J. and K. Virtanen. 2008. Game Theoretic Approach to Air Combat Simulation Model. Submitted for publication.
» Virtanen, K., T. Raivio, and R.P. Hämäläinen. 1999. Decision theoretical approach to pilot simulation. Journal of Aircraft 26 (4):632-641.