Optimisation de fonctions partiellement
Transcription
Optimisation de fonctions partiellement
Introduction Formalism Solve Example Conclusion Optimisation de fonctions partiellement modélisées avec requêtes contraintes à la vérité de terrain Frédéric Dambreville Lab-STICC UMR CNRS 6285, ENSTA Bretagne DGA MI, Bruz France Email: [email protected] PFMC 2013 – BREST – October 2013 Frédéric Dambreville Optimisation de fonctions partiellement modélisées 1 Introduction Formalism Solve Example Conclusion Introduction Main problem of interest Optimize a partially modelled function Assess the computed solution by actual evaluation Actual evaluation implies costly subprocess calls ← Subprocess may be a local optimization, a simulation. . . Each actual evaluation improves the function model Frédéric Dambreville Optimisation de fonctions partiellement modélisées 2 Introduction Formalism Solve Example Conclusion Introduction Main problem of interest Optimize a partially modelled function Assess the computed solution by actual evaluation How to combine model-based optimization and costly assessment ? Frédéric Dambreville Optimisation de fonctions partiellement modélisées 2 Introduction Formalism Solve Example Conclusion Introduction Main problem of interest Optimize a partially modelled function Assess the computed solution by actual evaluation How to combine model-based optimization and costly assessment ? Application context : Plan an Information collection by a team of sensors → Plan Manager only knows partial models ⇒ Optimize sensor-to-mission allocation and sensor route Full model for specific sensor is known by local teams → Plan sensor locally, but cannot plan a team of sensor ⇒ Assessment of specific sensor plan Frédéric Dambreville Optimisation de fonctions partiellement modélisées 2 Introduction Formalism Solve Example Conclusion Introduction Main problem of interest Optimize a partially modelled function Assess the computed solution by actual evaluation How to combine model-based optimization and costly assessment ? Application context : Plan an Information collection by a team of sensors → Plan Manager only knows partial models ⇒ Optimize sensor-to-mission allocation and sensor route Full model for specific sensor is known by local teams → Plan sensor locally, but cannot plan a team of sensor ⇒ Assessment of specific sensor plan Issue : Cooperation cost between manager and local teams Frédéric Dambreville Optimisation de fonctions partiellement modélisées 2 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y A prior law pµ of µ is known by the plan manager The true map µ b is unknown by the plan manager The true map µ b is known by the Local teams Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y valid trip τ [k] = τ [k|ik ] = y [k]x[m1k ] · · · x[mikk ]y [k] of k Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y Cost of valid trip τ [k] = τ [k|ik ] = y [k]x[m1k ] · · · x[mikk ]y [k] of k is C (τ [k]; µ) = c[y [k], x[m1 ]; µ] + c[x[m1 ], x[m2 ]; µ] · · · + c[x[mj ], y [k]; µ] Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y Cost of valid trip τ [k] = τ [k|ik ] = y [k]x[m1k ] · · · x[mikk ]y [k] of k is C (τ [k]; µ) = c[y [k], x[m1 ]; µ] + c[x[m1 ], x[m2 ]; µ] · · · + c[x[mj ], y [k]; µ] Optimize a mission assignment k 7→ τ [k] which : P maximize : G [τ ; µ] − ǫ k=1:K C (τ [k]; µ) where 0 < ǫ ≪ 1 and G [τ ; µ] = card ({m = 1 : M /∃k = 1 : K , m ∈ τ [k] }) in compliance with : C (τ [k]; µ) ≤ γ[k] for k = 1 : K Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Sensor-to-mission, simplified scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y Cost of valid trip τ [k] = τ [k|ik ] = y [k]x[m1k ] · · · x[mikk ]y [k] of k is C (τ [k]; µ) = c[y [k], x[m1 ]; µ] + c[x[m1 ], x[m2 ]; µ] · · · + c[x[mj ], y [k]; µ] Optimize a mission assignment k 7→ τ [k] which : P maximize : G [τ ; µ] − ǫ k=1:K C (τ [k]; µ) where 0 < ǫ ≪ 1 Solution assessed by local teams : C (τ [k]; µ b) > γ[k] ⇒ truncated to maximal valid trip τb[k] = τ [k|bik ] where C (τ [k|bik ]; µ b) ≤ γ[k] < C (τ [k|bik + 1]; µ b) P Solution is evaluated by G [b τ; µ b] − ǫ k=1:K C (b τ; µ b) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 3 Introduction Formalism Solve Example Conclusion Maximize a function with model noise Is given : f : (x, ν) ∈ X × N 7→ f (x, ν) where : x ∈ X is a parameter to be optimized ν ∈ N is a model noise pν ∈ P (N) is a known probabilistic noise prior and an unknown actual model noise : νb ∈ N is the unknown actual value of the model noise Frédéric Dambreville Optimisation de fonctions partiellement modélisées 4 Introduction Formalism Solve Example Conclusion Maximize a function with model noise Is given : f : (x, ν) ∈ X × N 7→ f (x, ν) where : x ∈ X is a parameter to be optimized ν ∈ N is a model noise pν ∈ P (N) is a known probabilistic noise prior and an unknown actual model noise : νb ∈ N is the unknown actual value of the model noise Issue : How to optimize f (·|b ν ) when each evaluation of f (·|b ν ) is obtained by a costly request ? Frédéric Dambreville Optimisation de fonctions partiellement modélisées 4 Introduction Formalism Solve Example Conclusion Reference work Efficient Global Optimization by Experiments Design [Jones, Schonlau, Welch, Efficient Global Optimization of Expensive Black-Box Functions, JoGO’98] Inspiration : Experiments Design Objective : maximize model-noised function f (x, ν) = g (x) + ν(x) Where ν is a functional Gaussian noise prior Correlation of ν(x1 ) and ν(x2 ) depends on d(x1 , x2 ) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 5 Introduction Formalism Solve Example Conclusion Reference work Efficient Global Optimization by Experiments Design [Jones, Schonlau, Welch, Efficient Global Optimization of Expensive Black-Box Functions, JoGO’98] Inspiration : Experiments Design Objective : maximize model-noised function f (x, ν) = g (x) + ν(x) Where ν is a functional Gaussian noise prior Correlation of ν(x1 ) and ν(x2 ) depends on d(x1 , x2 ) Optimizing f on the basis of iterated experiments Each measure f (b xk ) of function f is costly Optimize each measure → enhance the knowledge of ν around optimal values of f Frédéric Dambreville Optimisation de fonctions partiellement modélisées 5 Introduction Formalism Solve Example Conclusion Reference work Efficient Global Optimization by Experiments Design [Jones, Schonlau, Welch, Efficient Global Optimization of Expensive Black-Box Functions, JoGO’98] Inspiration : Experiments Design Objective : maximize model-noised function f (x, ν) = g (x) + ν(x) Where ν is a functional Gaussian noise prior Correlation of ν(x1 ) and ν(x2 ) depends on d(x1 , x2 ) Method : iterated measures maximizing Expected Improvement b xk+1 ∈ arg max EI (x) , x xi ) f (b x1:k ) . with : EI (x) = E max 0, f (x) − max f (b i=1:k Additive Gaussian ⇒ computation of the conditional law Frédéric Dambreville Optimisation de fonctions partiellement modélisées 5 Introduction Formalism Solve Example Conclusion Reference work Efficient Global Optimization by Experiments Design [Jones, Schonlau, Welch, Efficient Global Optimization of Expensive Black-Box Functions, JoGO’98] Inspiration : Experiments Design Objective : maximize model-noised function f (x, ν) = g (x) + ν(x) Where ν is a functional Gaussian noise prior Correlation of ν(x1 ) and ν(x2 ) depends on d(x1 , x2 ) Method : iterated measures maximizing Expected Improvement b xk+1 ∈ arg max EI (x) , x xi ) f (b x1:k ) . with : EI (x) = E max 0, f (x) − max f (b i=1:k Additive Gaussian ⇒ computation of the conditional law Our concern : model noise is not additive Gaussian Our approach : rare event simulation of the conditional law Frédéric Dambreville Optimisation de fonctions partiellement modélisées 5 Introduction Formalism Solve Example Conclusion Generalization The EGO generalizes easily to our problem : [Expected Improvement Maximization (EIM)] [Purpose : Optimize actual criterion f (·, ν b)] 1 Set n = 0 , 2 Repeat : 1 Compute b xn+1 , the next candidate for an actual evaluation : b xn+1 ∈ arg max x∈X Z pν [n](ν)f [n](x , ν) dν , ν∈N where : 2 3 xk , ν) = b yk and f [n](x , ν) = max pν [n](ν) = pν ν ∀k = 1 : n, f (b Request the actual evaluation of b xn+1 : b yn+1 = f (b xn+1 , ν b) f (x , ν), max b yk k=1:n Set n ← n + 1 , y1:n ) is sufficient. x1:n , b until the convergence of (b [Output :] The sequence (b x1:n , b y1:n ) , The model noise estimation pν [n] . Issue : how to estimate the conditional probability ? Frédéric Dambreville Optimisation de fonctions partiellement modélisées 6 Introduction Formalism Solve Example Conclusion Rare event simulation & conditional law The conditional law is approximated by : xk , ν) = b yk = pν ν φ(ν) ≥ −ǫ) pν ν ∀k = 1 : n, f (b X xk , ν), ybk and 0 < ǫ ≪ 1 where :φ(ν) = − d f (b k Simulation of rare event φ−1 (] − ǫ, 0]) is used in order to estimate the conditional probability The cross-entropy method is considered De Boer, Kroese, Mannor, Rubinstein, A Tutorial on the Cross-Entropy Method for the simulations for the optimization Frédéric Dambreville Optimisation de fonctions partiellement modélisées 7 Introduction Formalism Solve Example Conclusion Cross Entropy for simulation [CE simulation] 1 2 [Purpose : Build an Importance sampler for the rare event φ−1 ([γ, +∞[)] Set t = 0 and θ0 = θ o such that p = π(·|θ0 ) Repeat : 1 Generate the samples ωti ∈ Ω, i ∈ {1 : Nt }, according to π(·|θt ) 2 Compute the evaluations φ(ωti ) of the samples 3 Compute the selective parameters : ρt [0] = 1 − αt and ρt [i ] = αt P Rt (φ(ωti )) i i =1:Nt Rt (φ(ωt )) × π(ωti |θ o ) π(ωti |θt ) , for all i ∈ {1 : Nt } (1) // Rt is a selection map ; eg. Rt selects a quantile of best solution 4 [Update] Update the importance sampler by maximizing the cross-entropy with the selected samples : ! Z X i θt+1 ∈ arg max ρt [0]π(ω|θt ) + ρt [i ]δ[ω = ωt ] log (π(ω|θ)) dω θ∈Θ ω∈Ω i =1:Nt 5 Set t ← t + 1 until the convergence is sufficient, ie. the importance sampling of φ−1 ([γ, +∞[) is sufficient p (ω) ω [Output :] The importance sampler π(·|θt ) and the likelihood ratio W (ω|θt ) = π(ω|θ t) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 8 (2) Introduction Formalism Solve Example Conclusion Cross Entropy for simulation [CE simulation] 1 2 [Purpose : Build an Importance sampler for the rare event φ−1 ([γ, +∞[)] Set t = 0 and θ0 = θ o such that p = π(·|θ0 ) Repeat : 1 Generate the samples ωti ∈ Ω, i ∈ {1 : Nt }, according to π(·|θt ) 2 Compute the evaluations φ(ωti ) of the samples 3 Compute the selective parameters : ρt [0] = 1 − αt and ρt [i ] = αt P Rt (φ(ωti )) i i =1:Nt Rt (φ(ωt )) × π(ωti |θ o ) π(ωti |θt ) , for all i ∈ {1 : Nt } (1) // Rt is a selection map ; eg. Rt selects a quantile of best solution 4 [Update] Update the importance sampler by maximizing the cross-entropy with the selected samples : ! Z X i θt+1 ∈ arg max ρt [0]π(ω|θt ) + ρt [i ]δ[ω = ωt ] log (π(ω|θ)) dω θ∈Θ ω∈Ω i =1:Nt 5 Set t ← t + 1 until the convergence is sufficient, ie. the importance sampling of φ−1 ([γ, +∞[) is sufficient p (ω) ω [Output :] The importance sampler π(·|θt ) and the likelihood ratio W (ω|θt ) = π(ω|θ t) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 8 (2) Introduction Formalism Solve Example Conclusion Cross Entropy for optimization [CE optimization] [Purpose : Find a maximizer of φ] 1 Set t = 0 and θ0 = θ o 2 Repeat : 1 2 3 Generate the samples ωti ∈ Ω, i ∈ {1 : Nt }, according to π(·|θt ) Compute the evaluations φ(ωti ) of the samples Compute the selective parameters : Rt (φ(ωti )) , i i =1:Nt Rt (φ(ωt )) ρt [0] = 1 − αt and ρt [i ] = αt P for all i ∈ {1 : Nt } (3) // Rt is a selection map ; eg. Rt selects a quantile of best solution 4 [Update] Update the importance sampler by maximizing the cross-entropy with the selected samples : θt+1 ∈ arg max θ∈Θ Z ρt [0]π(ω|θt ) + ω∈Ω X i =1:Nt i ! ρt [i ]δ[ω = ωt ] log (π(ω|θ)) dω 5 Set t ← t + 1 until the convergence is sufficient [Output :] An estimation of the maximizers by means of the importance sampler π(·|θt ) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 9 (4) Introduction Formalism Solve Example Conclusion Back to sensor-to-mission scenario Requests m = 1 : M located at x[m] Sensors k = 1 : K with autonomy γ[k] and initial location y [k] A noisy map µ implying cost c[x, y ; µ] for moving from x to y Cost C (τ [k]; µ) = c[y [k], x[m1 ]; µ] + · · · + c[x[mj ], y [k]; µ] Optimize a mission assignment k 7→ τ [k] which : P maximize : G [τ ; µ] − ǫ k=1:K C (τ [k]; µ) where 0 < ǫ ≪ 1 Solution assessed by local teams : C (τ [k]; µ b) > γ[k] ⇒ truncated to trip τb[k] = τ [k|bik ] Solution is evaluated by G [b τ; µ b] − ǫ P k=1:K C (b τ; µ b) Frédéric Dambreville Optimisation de fonctions partiellement modélisées 10 Introduction Formalism Solve Example Conclusion Setings Sensors : (1, 1), (1, 1), (9, 1), (9, 1), (5, 1) Autonomy : 1, 2, 1, 2, 2 Missions : 20 missions chosen uniformly on [1, 10] × [1, 10] Map of threats : Theoretical threat position, with noise νi ∼ N(0, diag(2, 2)) : µ = (2, 3) + ν1 , (5, 4) + ν2 , (4, 7) + ν3 , (2, 3) + ν4 } Actual threat position : µ b = (1, 1), (4, 6), (3, 7), (1, 4)} Frédéric Dambreville Optimisation de fonctions partiellement modélisées 11 Introduction Formalism Solve Example Conclusion Setings Sensors : (1, 1), (1, 1), (9, 1), (9, 1), (5, 1) Autonomy : 1, 2, 1, 2, 2 Missions : 20 missions chosen uniformly on [1, 10] × [1, 10] Map of threats : Theoretical threat position, with noise νi ∼ N(0, diag(2, 2)) : µ = (2, 3) + ν1 , (5, 4) + ν2 , (4, 7) + ν3 , (2, 3) + ν4 } Actual threat position : µ b = (1, 1), (4, 6), (3, 7), (1, 4)} Inferred cost : Local cost c[x] = 1/(1 + d(x, T )2 ) decrease with dist. to threats R Cost of a path x → y is C [x, y ; µ] = ω∈[x,y ] c[ω] dω NB : complex path costs are possible, by choosing optimal paths Frédéric Dambreville Optimisation de fonctions partiellement modélisées 11 Introduction Formalism Solve Example Conclusion Sampling family for simulation and optimization The map : vector → Gaussian sampling laws The plan : The choice of the sensor-to-mission assignment and the choice of a mission rank is encoded by vector : ⇒ The plan sampling is obtained by sampling a vector by a Gaussian law and by discretizing it into a plan The CE-update of Gaussian family is a well known computation → empirical means and covariance Frédéric Dambreville Optimisation de fonctions partiellement modélisées 12 Introduction Formalism Solve Example Conclusion A sequence of run Optimization based on the true : iter R opt samp 1 20 40 60 80 100 120 11.6 13.3 14.3 14.8 15.1 15.5 16 100 2K 4K 6K 8K 10K 12K 140 160 180 200 16.6 17.3 17.7 17.9 14K 16K 18K 20K A sampled solution : 0 7→ {3} , 1 7→ {2, 10, 12, 16} , 2 7→ {4, 0, 18} , 3 7→ {13, 6, 5, 15, 7, 8} , 4 7→ {17, 1, 9, 19, 14, 11} Runs of the EGO : 0.8 2.6 ǫ NaN 0.07 0.96 0.72 y 15.7 17 17.7 17.7 17.1 NaN b y 14.9 13.9 11.9 15.92 15.93 NaN Quality of the conditional estimation ← ǫ Difficulty : multi-modal conditional law Vs Gaussian sampler Frédéric Dambreville Optimisation de fonctions partiellement modélisées 13 Introduction Formalism Solve Example Conclusion Conclusion Main Problem : Optimization with request to costly subprocesses Related work : Efficient Global Optimization (EGO) [expected improvement maximization algorithm] Frédéric Dambreville Optimisation de fonctions partiellement modélisées 14 Introduction Formalism Solve Example Conclusion Conclusion Main Problem : Optimization with request to costly subprocesses Related work : Efficient Global Optimization (EGO) [expected improvement maximization algorithm] Generalization proposed to non additive Gaussian noise CE method used for simulating CE method used for optimizing Frédéric Dambreville Optimisation de fonctions partiellement modélisées 14 Introduction Formalism Solve Example Conclusion Conclusion Main Problem : Optimization with request to costly subprocesses Related work : Efficient Global Optimization (EGO) [expected improvement maximization algorithm] Generalization proposed to non additive Gaussian noise CE method used for simulating CE method used for optimizing Perspectives Multi-modal law for conditional sampling Performance comparison with pure EGO Comparison with general POMDP approaches Frédéric Dambreville Optimisation de fonctions partiellement modélisées 14 Introduction Formalism Solve Example Conclusion Conclusion Main Problem : Optimization with request to costly subprocesses Related work : Efficient Global Optimization (EGO) [expected improvement maximization algorithm] Generalization proposed to non additive Gaussian noise CE method used for simulating CE method used for optimizing Perspectives Multi-modal law for conditional sampling Performance comparison with pure EGO Comparison with general POMDP approaches Questions ? Frédéric Dambreville Optimisation de fonctions partiellement modélisées 14