Data-driven distributionally robust LQR
with multiplicative noise
Peter Coppens, Mathijs Schuurmans, Panagiotis Patrinos
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Control in uncertain environments model = reality generalize model ≠ reality safety conservative performance
Uncertain systems financial (bio-)chemical climate control power distribution robotics/autonomous vehicles
Learning to generalize
classical controlidentification and controller design are kept seperate.
Data identification Model synthesis Controller
Learning to generalize
coarse id controlexplicitly include ambiguity in the model throughout the design1.
Data identification Model synthesis Controller
Learning to generalize
coarse id controlexplicitly include ambiguity in the model throughout the design1.
Data identification Model synthesis Controller
Learning to generalize
coarse id controlexplicitly include ambiguity in the model throughout the design1.
Data identification Model synthesis Controller
Robustify — Stochastic and robust optimization • Robust optimization (𝑤 ∈ 𝒲) min 𝑥 max𝑤∈𝒲 𝑓(𝑥,𝑤) s. t. 𝑔(𝑥,𝑤) ≤ 0,∀𝑤 ∈ 𝒲 • Stochastic optimization (𝑤 ∼ IP𝑤) min 𝑥 IE[𝑓(𝑥,𝑤)] s. t. IP[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿
Robustify — Stochastic and robust optimization
• Robust optimization (𝑤 ∈ 𝒲)
min
𝑥 max𝑤∈𝒲 𝑓(𝑥,𝑤)
s. t. 𝑔(𝑥,𝑤) ≤ 0,∀𝑤 ∈ 𝒲
strong guarantees, conservative
• Stochastic optimization (𝑤 ∼ IP𝑤)
min
𝑥 IE[𝑓(𝑥,𝑤)]
s. t. IP[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿
Robustify — Stochastic and robust optimization • Robust optimization (𝑤 ∈ 𝒲) min 𝑥 max𝑤∈𝒲 𝑓(𝑥,𝑤) s. t. 𝑔(𝑥,𝑤) ≤ 0,∀𝑤 ∈ 𝒲 • Stochastic optimization (𝑤 ∼ IP𝑤) min 𝑥 IE[𝑓(𝑥,𝑤)] s. t. IP[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿
Robustify — Stochastic and robust optimization • Robust optimization (𝑤 ∈ 𝒲) min 𝑥 max𝑤∈𝒲 𝑓(𝑥,𝑤) s. t. 𝑔(𝑥,𝑤) ≤ 0,∀𝑤 ∈ 𝒲 • Stochastic optimization (𝑤 ∼ IP𝑤) min 𝑥 IE[𝑓(𝑥,𝑤)] s. t. IP[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿 less guarantees, optimal
Robustify — Distributionally robust optimization • Distributionally robust (𝜉 ∼ IP𝑤,IP𝑤∈ 𝒜) min 𝑥 IPmax𝑤∈𝒜 IE𝑤[𝑓(𝑥,𝑤)] s. t. IP𝑤[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿,∀IP𝑤∈ 𝒜
conservatism decreases with data akin to automatic regularization2 already applied in control before3-5
Robustify — Distributionally robust optimization • Distributionally robust (𝜉 ∼ IP𝑤,IP𝑤∈ 𝒜) min 𝑥 IPmax𝑤∈𝒜 IE𝑤[𝑓(𝑥,𝑤)] s. t. IP𝑤[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿,∀IP𝑤∈ 𝒜
conservatism decreases with data
akin to automatic regularization2 already applied in control before3-5
Robustify — Distributionally robust optimization • Distributionally robust (𝜉 ∼ IP𝑤,IP𝑤∈ 𝒜) min 𝑥 IPmax𝑤∈𝒜 IE𝑤[𝑓(𝑥,𝑤)] s. t. IP𝑤[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿,∀IP𝑤∈ 𝒜
conservatism decreases with data akin to automatic regularization2
Robustify — Distributionally robust optimization • Distributionally robust (𝜉 ∼ IP𝑤,IP𝑤∈ 𝒜) min 𝑥 IPmax𝑤∈𝒜 IE𝑤[𝑓(𝑥,𝑤)] s. t. IP𝑤[𝑔(𝑥,𝑤) ≤ 0] ≥ 1 − 𝛿,∀IP𝑤∈ 𝒜
conservatism decreases with data akin to automatic regularization2 already applied in control before3-5
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given 𝑤(𝑖)𝑘 𝑤(𝑖)𝑘
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf 𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]} IE [𝑥⊤ +𝑃 𝑥+ ∣ (𝑥, 𝑢)] = (A0𝑥 + B0𝑢) ⊤ ([1 𝜇 ⊤ 𝜇 Σ + 𝜇𝜇⊤] ⊗ 𝑃) (A0𝑥 + B0𝑢) , whereA0= [𝐴⊤0 𝐴⊤1 … 𝐴⊤𝑛𝑤] ⊤ andB0= [𝐵⊤0 𝐵1⊤ … 𝐵𝑛⊤𝑤] ⊤
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation1-3
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant true optimum
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation1-3
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞
• performant true optimum • only depends on 𝜇 and Σ
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation1-3
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant true optimum
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation1-3
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant true optimum
DR-LQR — The nominal problem minimize 𝑢0,𝑢1,… IE [ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = (𝐴0+ ∑𝑛𝑤 𝑖=1𝐴𝑖𝑤 (𝑖) 𝑘 ) 𝑥𝑘+ (𝐵0+ ∑ 𝑛𝑤 𝑖=0𝐵𝑖𝑤 (𝑖) 𝑘 ) 𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + IE [𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation1-3
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant true optimum
DR-LQR — Learning to generalize
Data identification Model Controller synthesis
DR-LQR — Learning to generalize
Data identification Model Controller synthesis
quantify robustify
assumptions about the data:
• access to samples {𝑤𝑖}𝑀𝑖=1 • Σ−1/2(𝑤 − 𝜇) is sub-Gaussian • 𝜇̂= 1 𝑀∑ 𝑀 𝑖=1𝑤𝑖 • Σ̂ = 𝑀1 ∑𝑖=1𝑀 (𝑤𝑖− ̂𝜇)(𝑤𝑖− ̂𝜇)⊤
DR-LQR — Learning to generalize
Data identification Model Controller synthesis
quantify robustify
assumptions about the data:
• access to samples {𝑤𝑖}𝑀𝑖=1 • Σ−1/2(𝑤 − 𝜇) is sub-Gaussian1 • 𝜇̂= 1 𝑀∑ 𝑀 𝑖=1𝑤𝑖 • Σ̂ = 𝑀1 ∑𝑖=1𝑀 (𝑤𝑖− ̂𝜇)(𝑤𝑖− ̂𝜇)⊤
DR-LQR — Learning to generalize
Data identification Model Controller synthesis
quantify robustify
assumptions about the data:
• access to samples {𝑤𝑖}𝑀𝑖=1 • Σ−1/2(𝑤 − 𝜇) is sub-Gaussian1 • 𝜇̂= 1 𝑀∑ 𝑀 𝑖=1𝑤𝑖 • Σ̂ = 𝑀1 ∑𝑖=1𝑀 (𝑤𝑖− ̂𝜇)(𝑤𝑖− ̂𝜇)⊤
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤
Σ⪯ Σ̂ }
𝑟𝜇(𝛽) 𝑟Σ(𝛽)
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch thatIP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
Σ⪯ 𝑟Σ(𝛽)Σ̂ }
• radii decrease with 1/√# samples
• no knowledge of true distribution required • sufficient samples required
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
Σ⪯ 𝑟Σ(𝛽)Σ̂ }
• radii decrease with 1/√# samples
• no knowledge of true distribution required
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
Σ⪯ 𝑟Σ(𝛽)Σ̂ }
• radii decrease with 1/√# samples
• no knowledge of true distribution required • sufficient samples required
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽
𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
DR-LQR — The ambiguity set
quantization: determine 𝑟𝜇and 𝑟Σsuch that IP [IP⋆𝑤 ∈ 𝒜] ≥ 1 − 𝛽 𝒜 ∶= {IP𝑤∣ (𝜇−𝜇)̂
⊤Σ̂−1(𝜇−𝜇̂) ≤ 𝑟 𝜇(𝛽)
Σ⪯ 𝑟Σ(𝛽)Σ̂ }
synthesis: find the distributionally robust controller
min 𝑢0,𝑢1,…IPmax𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,… IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given max Σ⪯𝑟ΣΣ̂
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) +Σ⪯𝑟max ΣΣ̂
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf 𝑢 {ℓ(𝑥, 𝑢) + maxΣ⪯𝑟 ΣΣ̂ IE𝑤[𝑉 (𝑥+) ∣ (𝑥, 𝑢)]} max Σ⪯𝑟ΣΣ̂ IE𝑤[𝑥⊤+𝑃 𝑥+ ∣ (𝑥, 𝑢)] = (A0𝑥 + B0𝑢) ⊤ ([1 𝜇 ⊤ 𝜇 𝑟ΣΣ̂+ 𝜇𝜇⊤] ⊗ 𝑃) (A0𝑥 + B0𝑢)
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + maxΣ⪯𝑟 ΣΣ̂
IE𝑤[𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + maxΣ⪯𝑟 ΣΣ̂
IE𝑤[𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing w.h.p. IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞
• performant converges to true optimum • only depends on 𝜇 and𝑟ΣΣ̂
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + maxΣ⪯𝑟 ΣΣ̂
IE𝑤[𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing w.h.p. IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant converges to true optimum
DR-LQR — Uncertain covariance min 𝑢0,𝑢1,…Σ⪯𝑟max ΣΣ̂ IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solution is still quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) + maxΣ⪯𝑟 ΣΣ̂
IE𝑤[𝑉 (𝑥+) ∣ (𝑥, 𝑢)]}
generalized Riccati equation
properties of optimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing w.h.p. IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant converges to true optimum • only depends on 𝜇 and𝑟ΣΣ̂
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given max IP𝑤∈𝒜
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:solutionno longer quadratic 𝑉 (𝑥) = inf
𝑢 {ℓ(𝑥, 𝑢) +IPmax𝑤∈𝒜
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:approximate solution1
minimize 𝑃 𝑥 ⊤ 0𝑃 𝑥0 subj. to 𝑥⊤𝑃 𝑥 ≥ min 𝑢 {ℓ(𝑥, 𝑢) + maxIP𝑤∈𝒜IE𝑤[𝑥 ⊤ +𝑃 𝑥+]} , ∀𝑥
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:approximate solution1
minimize 𝑃 𝑥 ⊤ 0𝑃 𝑥0 subj. to 𝑥⊤𝑃 𝑥 ≥ min 𝑢 {ℓ(𝑥, 𝑢) + maxIP𝑤∈𝒜IE𝑤[𝑥 ⊤ +𝑃 𝑥+]} , ∀𝑥 solve approximate SDP2
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:approximatesolution1
minimize 𝑃 𝑥 ⊤ 0𝑃 𝑥0 subj. to 𝑥⊤𝑃 𝑥 ≥ min 𝑢 {ℓ(𝑥, 𝑢) + maxIP𝑤∈𝒜 IE𝑤[𝑥⊤ +𝑃 𝑥+]} , ∀𝑥 solveapproximateSDP2
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:approximate solution1
minimize 𝑃 𝑥 ⊤ 0𝑃 𝑥0 subj. to 𝑥⊤𝑃 𝑥 ≥ min 𝑢 {ℓ(𝑥, 𝑢) + maxIP𝑤∈𝒜IE𝑤[𝑥 ⊤ +𝑃 𝑥+]} , ∀𝑥 solve approximate SDP2
properties of approximateoptimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing w.h.p. IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞
DR-LQR — Full uncertainty min 𝑢0,𝑢1,… max IP𝑤∈𝒜 IE𝑤[ ∞ ∑ 𝑘=0 𝑥⊤ 𝑘𝑄𝑥𝑘+ 𝑢⊤𝑘𝑅𝑢𝑘] subj. to 𝑥𝑘+1 = 𝐴(𝑤𝑘)𝑥𝑘+ 𝐵(𝑤𝑘)𝑢𝑘, ∀𝑘 ∈ IN 𝑥0given
Bellman equation:approximate solution1
minimize 𝑃 𝑥 ⊤ 0𝑃 𝑥0 subj. to 𝑥⊤𝑃 𝑥 ≥ min 𝑢 {ℓ(𝑥, 𝑢) + maxIP𝑤∈𝒜IE𝑤[𝑥 ⊤ +𝑃 𝑥+]} , ∀𝑥 solve approximate SDP2
properties of approximateoptimal controller:
• linear 𝑢 = 𝐾𝑥 (solve SDP)
• safe mean-square stabilizing w.h.p. IE[𝑥⊤
𝑘𝑥𝑘] → 0 as 𝑘 → ∞ • performant converges to true optimum
DR-LQR — Numerical result 103 104 105 106 10−6 10−5 10−4 10−3 10−2 # samples R el. Subop t. uncertain covariance full uncertainty
DR-LQR — Numerical result 103 104 105 106 10−6 10−5 10−4 10−3 10−2 # samples R el. Subop t. uncertain covariance full uncertainty 1/# samples
DR-LQR — Future work
• what to do when not enough samples?
• what to do when only state measurements? • constraints, state observers, …
DR-LQR — Future work
• what to do when not enough samples? • what to do when only state measurements?
DR-LQR — Future work
• what to do when not enough samples? • what to do when only state measurements? • constraints, state observers, …
References i
Abadeh, Soroosh Shafieezadeh, Peyman Mohajerin Esfahani, and Daniel Kuhn.
“Distributionally Robust Logistic Regression”.In: Advances in Neural Information Processing Systems 28. 2015, pp. 1576–1584.
Abadeh, Soroosh Shafieezadeh et al.“Wasserstein distributionally robust Kalman filtering”.In: Advances in Neural Information Processing Systems. 2018, pp. 8474–8483. Ben-Tal, Aharon, Laurent El Ghaoui, and Arkadi Nemirovski.“Robustness”.In: Handbook
of Semidefinite Programming. Boston, MA: Springer US, 2000, pp. 139–162.
Boyd, Stephen et al.Linear Matrix Inequalities in System and Control Theory.Society for Industrial and Applied Mathematics, 1994.
Coppens, P. and P. Patrinos.Sample Complexity of Data-Driven Stochastic LQR with Multiplicative Uncertainty.2020. arXiv: 2005.12167.
References ii
Costa, Oswaldo L.V. and Carlos S. Kubrusly.“Quadratic optimal control for discrete-time infinite-dimensional stochastic bilinear systems”.In: IMA Journal of Mathematical Control and Information 14.4 (1997), pp. 385–399.
Delage, Erick and Yinyu Ye.“Distributionally robust optimization under moment uncertainty with application to data-driven problems”.In: Operations Research 58.3 (2010), pp. 595–612.
Gravell, Benjamin, Peyman Mohajerin Esfahani, and Tyler Summers.“Learning robust control for LQR systems with multiplicative noise via policy gradient”.In: arXiv preprint arXiv:1905.13547 (2019). arXiv: 1905.13547.
Hsu, Daniel, Sham M. Kakade, and Tong Zhang.“A tail inequality for quadratic forms of subgaussian random vectors”.In: Electronic Communications in Probability 17.52 (2012), pp. 1–6.
References iii
Hsu, Daniel, Sham M. Kakade, and Tong Zhang.“Tail inequalities for sums of random matrices that depend on the intrinsic dimension”.In: Electronic Communications in Probability 17.52 (2012), pp. 1–13.
Rahimian, H. and S. Mehrotra.Distributionally Robust Optimization: A Review.2019. arXiv: 1908.05659.
Recht, Benjamin.“A Tour of Reinforcement Learning: The View from Continuous Control”.In: Annual Review of Control, Robotics, and Autonomous Systems 2.1 (2019), pp. 253–279.
Schuurmans, Mathijs, Pantelis Sopasakis, and Panagiotis Patrinos.“Safe
Learning-Based Control of Stochastic Jump Linear Systems: a Distributionally Robust Approach”.In: 2019 IEEE Conference on Decision and Control (CDC). IEEE. 2019,
References iv
Vershynin, Roman.High-Dimensional Probability: An Introduction with Applications in Data Science.2018.
Wainwright, Martin J.High-Dimensional Statistics.Cambridge University Press, 2019. Wonham, W. Murray.“Optimal Stationary Control of a Linear System with
State-Dependent Noise”.In: SIAM Journal on Control 5.3 (1967), pp. 486–500. Yang, Insoon.Wasserstein distributionally robust stochastic control: A data-driven
Assets
Images‘control in uncertain environments’: https://flic.kr/p/<e6tYCK,939fWN,2iUoC88> Images‘uncertain systems’:
https://flic.kr/p/<k283C7,xz8h9,2hRaa8G,r3LtVQ,azfYAW,n9mn42> Matlab code available on request
Quantify – Sub-Gaussian
Definition (Sub-Gaussian)
Random variable 𝑋 is sub-Gaussian1if, for some 𝜎 ≥ 0, IE[𝑒𝜆(𝑋−IE[𝑋])] ≤ 𝑒𝜎2𝜆2/2
, ∀𝜆 ∈ IR
examples2:
• Gaussian random variables 𝜎2= IE[(𝑋 − IE[𝑋])2],
• Bernoulli random variable 𝜎 = 1/√log(2),
Quantify – Sub-Gaussian
Definition (Sub-Gaussian)
Random variable 𝑋 is sub-Gaussian1if, for some 𝜎 ≥ 0, IE[𝑒𝜆(𝑋−IE[𝑋])] ≤ 𝑒𝜎2𝜆2/2
, ∀𝜆 ∈ IR
examples2:
• Gaussian random variables 𝜎2= IE[(𝑋 − IE[𝑋])2],
• Bernoulli random variable 𝜎 = 1/√log(2),
• random variables with bounded support 𝜎 = ess sup |𝑋|/√log(2). Sub-Gaussian vector𝑍: for every 𝑢 ∈ 𝒮𝑑we have 𝑢⊤𝑍 sub-Gaussian.
Quantify – Sub-Gaussian
Definition (Sub-Gaussian)
Random variable 𝑋 is sub-Gaussian1if, for some 𝜎 ≥ 0, IE[𝑒𝜆(𝑋−IE[𝑋])] ≤ 𝑒𝜎2𝜆2/2
, ∀𝜆 ∈ IR
examples2:
• Gaussian random variables 𝜎2= IE[(𝑋 − IE[𝑋])2],
• Bernoulli random variable 𝜎 = 1/√log(2),
• random variables with bounded support 𝜎 = ess sup |𝑋|/√log(2). Sub-Exponential: |𝜆| < 1/𝛼.
Quantify — Sum of random vectors/matrices
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟𝜇,Σ⪯𝑟ΣΣ̂} ,
Quantify — Sum of random vectors/matrices
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: IE[𝑤𝐼] = 0, IE[𝑤𝐼𝑤⊤𝐼] = 𝐼.
mean uncertainty: 𝑤𝐼∈ IR𝑑∼ subG(𝜎) 𝜇̂𝐼∼ subG𝑑(𝜎/
√ 𝑀). IP [‖𝜇̂𝐼‖2≤ 𝜎
2
𝑀(𝑛𝑤+ 2√𝑛𝑤log (1/𝛽) + 2log(1/𝛽))] ≥ 1 − 𝛽
Daniel Hsu, Sham M. Kakade, and Tong Zhang. “A tail inequality for quadratic forms of subgaussian random vectors”. In: Electronic Commu-nications in Probability 17.52 (2012), pp. 1–6
Quantify — Sum of random vectors/matrices
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: IE[𝑤𝐼] = 0, IE[𝑤𝐼𝑤⊤𝐼] = 𝐼. covariance uncertainty: 𝑤𝐼∈ IR𝑑∼ subG(𝜎).
IP [‖Σ̂𝐼− 𝐼‖2≤ 𝜎 2 1 − 2𝜖(√ 32𝑞(𝛽, 𝜖, 𝑛𝑤) 𝑀 + 2𝑞(𝛽, 𝜖, 𝑛𝑤) 𝑀 , )] ≥ 1 − 𝛽, where 𝑞(𝛽, 𝜖, 𝑛𝑤) = 𝑛𝑤log(1 +1/𝜖) + log(2/𝛽).
Daniel Hsu, Sham M. Kakade, and Tong Zhang. “Tail inequalities for sums of random matrices that depend on the intrinsic dimension”. In: Electronic Communications in Probability 17.52 (2012), pp. 1–13
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
𝑤𝐼= Σ−1/2𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
‖Σ̂𝐼− 𝐼‖2≤𝑡Σ
𝑤𝐼= Σ−1/2𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
(1 −𝑡Σ)𝐼 ⪯Σ̂𝐼 ⪯ (1 +𝑡Σ)𝐼
𝑤𝐼= Σ−1/2
𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
(1 −𝑡Σ)𝐼 ⪯Σ̂𝐼
𝑤𝐼= Σ−1/2𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
𝐼 ⪯ 1 (1 −𝑡Σ)Σ̂𝐼
𝑤𝐼= Σ−1/2𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0. Σ1/2𝐼Σ1/2⪯ 1 (1 −𝑡Σ)Σ 1/2Σ̂ 𝐼Σ 1/2 𝑤𝐼= Σ−1/2 𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0. Σ1/2𝐼Σ1/2⪯ 1 (1 −𝑡Σ)Σ 1/2Σ̂ 𝐼Σ 1/2 𝑤𝐼= Σ−1/2 𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽)
bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
illustration: for 𝜇 = 0.
Σ ⪯ 1
(1 −𝑡Σ)Σ̂
𝑤𝐼= Σ−1/2𝑤
Erick Delage and Yinyu Ye. “Distributionally robust optimization under moment uncertainty with application to data-driven problems”. In: Op-erations Research 58.3 (2010), pp. 595–612
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽) bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)).
Quantify — Final ambiguity set
𝒜 = {IP𝑤 ∣ (𝜇̂−𝜇)Σ̂−1(𝜇̂−𝜇) ≤𝑟
𝜇,Σ⪯𝑟ΣΣ̂} ,
first consider isotropic case: ‖𝜇̂𝐼‖2≤𝑡𝜇(𝛽) and ‖Σ̂𝐼− 𝐼‖2≤𝑡Σ(𝛽) bound:
𝑟Σ(𝛽) = 1/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)) and 𝑟𝜇(𝛽) =𝑡𝜇(𝛽/2)/(1−𝑡Σ(𝛽/2)−𝑡𝜇(𝛽/2)). only valid after 𝑀 sufficiently large (𝑡Σ(𝛽/2) +𝑡𝜇(𝛽/2) < 1)!
SDPs — Nominal problem
solutionexists when system is mean square stabilizable. • Riccati equation:
𝑃 = 𝑄 + 𝐹 (𝑃 ) − 𝐻(𝑃 )⊤(𝑅 + 𝐺(𝑃 ))−1𝐻(𝑃 )
• optimal controller:
SDPs — Nominal problem
solutionexists when system is mean square stabilizable. • Riccati equation: 𝑃 = 𝑄 + 𝐹 (𝑃 ) − 𝐻(𝑃 )⊤(𝑅 + 𝐺(𝑃 ))−1𝐻(𝑃 ) • optimal controller: 𝐾 = −(𝑅 + 𝐺(𝑃 ))−1𝐻(𝑃 ) Semidefinite program: minimize 𝑃 − Tr 𝑃 subj. to [𝑄 − 𝑃 + 𝐹 (𝑃 ) 𝐻(𝑃 ) ⊤ 𝐻(𝑃 ) 𝑅 + 𝐺(𝑃 )] ⪰ 0, 𝑃 ⪰ 0.
SDPs — Unknown covariance solution • Riccati equation: 𝑃 = 𝑄 + 𝐹 (𝑃 ) − 𝐻(𝑃 )⊤(𝑅 + 𝐺(𝑃 ))−1𝐻(𝑃 ) • optimal controller: 𝐾 = −(𝑅 + 𝐺(𝑃 ))−1𝐻(𝑃 ) Semidefinite program: only change covariance
minimize 𝑃 − Tr 𝑃 subj. to [𝑄 − 𝑃 +𝐹(𝑃 ) 𝐻(𝑃 ) ⊤ 𝐻(𝑃 ) 𝑅 +𝐺(𝑃 )] ⪰ 0, 𝑃 ⪰ 0.
SDPs — Full uncertainty
solution
• approximation of value function:
𝑉 (𝑥) = 𝑥⊤𝑊−1𝑥 • optimal controller: 𝐾 = 𝑉 𝑊−1 maximize 𝑊 ,𝑉 ,𝑆,𝐿 Tr𝑊 subj. to ⎡⎢ ⎣ 𝑆 𝑟𝜇𝐻1⊤𝑟𝜇𝐻⊤2 ⋯𝑟𝜇𝐻⊤𝑛𝑤 𝑟𝜇𝐻1 𝐿 𝑟𝜇𝐻2 𝐿 ⋮ ⋱ 𝑟𝜇𝐻𝑛𝑤 𝐿 ⎤ ⎥ ⎦ ⪰ 0, ⎡ ⎢ ⎢ ⎢ ⎣ 𝑊 −√2𝑆 (A𝑊 +B𝑉 )⊤ ( ̂𝐴𝑊 + ̂𝐵𝑉 )⊤𝑊⊤𝑄12 𝑉⊤𝑅12 A𝑊 +B𝑉 (𝑟ΣΣ̂)−1⊗𝑊 ̂ 𝐴𝑊 + ̂𝐵𝑉 𝑊 −√2𝐿 𝑄12𝑊 𝐼𝑛𝑥 𝑅12𝑉 𝐼𝑛𝑢 ⎤ ⎥ ⎥ ⎥ ⎦ ⪰ 0, ̂ 𝐴 = 𝐴 + ∑𝑛𝑤𝐴𝜇̂(𝑖),𝐵 = 𝐵̂ + ∑𝑛𝑤𝐵𝜇̂(𝑖), A = [𝐴⊤… 𝐴⊤ ]⊤, B = [𝐵⊤… 𝐵⊤ ]⊤and
SDPs — Full uncertainty
solution
• approximation of value function:
𝑉 (𝑥) = 𝑥⊤𝑊−1𝑥
• optimal controller:
𝐾 = 𝑉 𝑊−1
̂