Chance-constrained optimization with tight confidence bounds Mark Cannon University of Oxford 25 April 2017 1
Outline 1. Problem definition and motivation 2. Confidence bounds for randomised sample discarding 3. Posterior confidence bounds using empirical violation estimates 4. Randomized algorithm with prior and posterior confidence bounds 5. Conclusions and future directions 2
Chance-constrained optimization Chance constraints limit the probability of constraint violation in optimization problems with stochastic parameters Applications are numerous and diverse:... building control, chemical process design, finance, portfolio management, production planning, supply chain management, sustainable development, telecommunications networks... Long history of stochastic programming methods e.g. Charnes & Cooper, Management Sci., 1959 Prekopa, SIAM J. Control, 1966 but exact handling of probabilistic constraints is generally intractable This talk: approximate methods based on random samples 3
Motivation: Energy management in PHEVs Battery: charged before trip, empty at end of trip Electric motor: shifts engine load point to improve fuel e ciency or emissions Control problem: optimize power flows to maximize performance (e ciency or emissions) over the driving cycle 4
Motivation: Energy management in PHEVs Fuel P f Engine P eng Clutch Battery Circuit Electric P s Losses P b Motor P em + Gearbox Vehicle P drv Parallel configuration with common engine and electric motor speed P drv = P eng + P em! eng =! em =! (if!! stall ) I Engine fuel consumption: P f = P f (P eng,!) 5
Motivation: Energy management in PHEVs Fuel P f Engine P eng Clutch Battery Circuit Electric P s Losses P b Motor P em + Gearbox Vehicle P drv I Electric motor electrical power P b $ mechanical power P em : P b = P b (P em,!) I Battery stored energy flow rate P s $ P b : P s = P s (P b ) stored energy E $ P s : Ė = P s 5
Motivation: Energy management in PHEVs Nominal optimization, assuming known P drv (t),!(t) for t 2 [0,T]: minimize P eng ( ), P em ( ) 2[t,T ] Z T 0 P f (P eng ( ),!( )) d subject to Ė( ) = P s ( ), 2 [t, T ] E(t) =E now E(T ) E end with P drv = P eng + P em P b = P b (P em,!) P s = P s (P b ) 6
Motivation: Energy management in PHEVs P drv (t),!(t) depend on uncertain tra c flow and driver behaviour 0.6 Battery state of energy 0.58 0.56 0.54 0.52 0.5 200 400 600 800 1000 1200 1400 1600 1800 Time (sec) Introduce feedback by using a receding horizon implementation but how likely is it that we can meet constraints? 7
Motivation: Energy management in PHEVs Pdrv (t),!(t) depend on uncertain traffic flow and driver behaviour Real time traffic flow data 7
Motivation: Energy management in PHEVs Reformulate as a chance-constrained optimization, assuming known distributions for P drv (t),!(t): minimize P eng ( ), P em ( ) 2[t,T ] Z T 0 P f (P eng ( ),!( )) d subject to Ė( ) = P s ( ), 2 [t, T ] E(t) =E now P E(T ) <E end apple for a specified probability 8
Chance-constrained optimization Define the chance-constrained problem CCP : minimize x2x c > x subject to P f(x, ) > 0 apple. 2 R d stochastic parameter 2 [0, 1] maximum allowable violation probability Assume X R n compact and convex f(, ):R n! R convex, lower-semicontinuous for any 2 CCP is feasible, i.e. P{f(x, ) > 0} apple for some x 2X. 9
Chance-constrained optimization Computational di culties with chance constraints: 1. Prohibitive computation to determine feasibility of a given x Z P f(x, ) > 0 = 1 {f(x, )>0} ( ) F (d ) 2 Closed form expression only available in special cases N(0,I) Gaussian e.g. if f(x, )=b > 1+(d + D > ) > x a ne then P f(x, ) > 0 apple () 1 N (1 )kb + Dxk apple1 d> x In general multidimensional integration is needed (e.g. Monte Carlo methods) 10
Chance-constrained optimization Computational di culties with chance constraints: 2. Feasible set may be nonconvex even if f(, ) is convex for all 2 e.g. 1 N (1 ) kb + Dxk apple1 d> x is convex if apple 0.5 nonconvex if >0.5 x feasible x feasible <0.5 >0.5 10
Sample approximation Define the multisample! m = { (1),..., (m) } for i.i.d. (i) 2 Define the sample approximation of CCP as SP : minimize x2x!! m c > x subject to f(x, ) apple 0 for all 2!! q m q discarded constraints, m q n x (! ): optimal solution for x! : optimal solution for! 11
Sample approximation B Solutions of SP converge to the solution of CCP as m!1 if q = m(1 ) B Approximation accuracy characterized via bounds on: probability that solution of SP is feasible for CCP probability that optimal objective of SP exceeds that of CCP Luedtke & Ahmed, SIAM J. Optim., 2008 Calafiore, SIAM J. Optim., 2010 Campi & Garatti, J. Optim. Theory Appl., 2011 B SP can be formulated as a mixed integer program with a convex continuous relaxation 12
Example: minimal bounding circle Smallest circle containing a given probability mass CCP : minimize c,r R subject to P kc k >R apple SP : minimize c,r!! m R subject to kc kappler for all 2!! q e.g. =0.5, m =1000, q =500 R c CCP 1.386 (0, 0) SP 1.134 ( 0.003, 0.002) SP solved for one realisation! 1000 N(0,I) 13
Example: minimal bounding circle Bounds on confidence that SP solution, x (! ) with m = 1000, q = 500 is feasible for CCP ( ) from Calafiore, 2010 1 0.95-0.9 0.8 0.7 Confidence 0.6 0.5 0.4 0.3 0.2 0.95 0.05 =0.11-14 0.1 0.05-0 0.4 0.45 0.5 0.55 0.6 Violation probability ϵ =) with 90% probability: the violation probability of x (! ) lies between 0.47 and 0.58
Example: minimal bounding circle Smallest circle containing a given probability mass observations: SP computation grows rapidly with m Bounds on confidence of feasibility of x (! ) for CCP are not tight (unless q<mor probability distribution is a specific worst case) because SP solution has poor generalization properties e.g. m = 20, q = 10: 15 Randomized sample discarding improves generalization
Outline 1. Problem definition and motivation 2. Confidence bounds for randomised sample discarding 3. Posterior confidence bounds using empirical violation estimates 4. Randomized algorithm with prior and posterior confidence bounds 5. Conclusions and future directions 16
Level-q subsets of a multisample Define the sampled problem for given!! m as SP(!) : minimize x2x c > x subject to f(x, ) apple 0 for all 2! x (!): optimal solution of SP(!) Assumptions: (i). SP(!) is feasible for any!! m with probability (w.p.) 1 (ii). f x (!), 6= 0 for all 2! m \! w.p. 1 17
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 2 10 (! 20 ) 18
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 2 10 (! 20 ) 18
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 2 10 (! 20 ) 18
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 2 10 (! 20 ) 18
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 2 10 (! 20 ) 18
Level-q subsets of a multisample Define the set of level-q subsets of! m,forq apple m as q (! m )= n o!! m :! = q and f x (!), > 0 for all 2! m \! x (!) for! 2 q (! m ) is called a level-q solution e.g. minimal bounding circle problem, m = 20, q = 10:! 62 10 (! 20 ) 18
Essential constraint set Es(!) ={ (i1),..., (i k ) }! is an essential set of SP(!) if (i). x { (i1),..., (i k ) } = x (!), (ii). x(! \ ) 6= x (!) for all 2{ (i1),..., (i k ) }. e.g. minimal bounding circle problem CO Pi C PPPPP C C essential constraints 19
Support dimension Define: maximum support dimension: =esssup!2 m Es(!) m 1 minimum support dimension: =essinf!2 m Es(!) then apple dim(x) =n and 1 e.g. minimal bounding circle problem m Es(!) =2= 20
Support dimension Define: maximum support dimension: =esssup!2 m Es(!) m 1 minimum support dimension: =essinf!2 m Es(!) then apple dim(x) =n and 1 e.g. minimal bounding circle problem m Es(!) =3= 20
Confidence bounds for randomly selected level-q subsets Theorem 1: For any 2 [0, 1] and q 2 [,m], P m V x (!) apple! 2 q (! m ) (q ; m, 1 ). For any 2 [0, 1] and q 2 [,m], P m V x (!) apple! 2 q (! m ) apple (q ; m, 1 ). where violation probability: V (x) =P f(x, ) > 0 binomial c.d.f.: (n; N,p) = nx i=0 N p i (1 p) N i i 21
Confidence bounds for randomly selected level-q subsets Proof of Thm. 1 relies on... Lemma 1: For any v 2 [0, 1] and k 2 [, ], P m V x (! k ) apple v Es(! k ) = k = F (v) =v k (e.g. Calafiore, 2010) Proof (sketch): for all q 2 [k, m], P m! k =Es(! q) \ V x (! k ) = v Es(! k ) = k =(1 v) q k df (v) =) P m! k =Es(! q) Es(! k ) = k = Z 1 0 (1 v) q k df (v) 22 hence Z 1 0 (1 v) q k q 1 df (v) = Hausdor momentproblem k
Confidence bounds for randomly selected level-q subsets Proof of Thm. 1 relies on... Lemma 1: For any v 2 [0, 1] and k 2 [, ], P m V x (! k ) apple v Es(! k ) = k = F (v) =v k (e.g. Calafiore, 2010) Proof (sketch): for all q 2 [k, m], P m! k =Es(! q) \ V x (! k ) = v Es(! k ) = k =(1 v) q k df (v) =) P m! k =Es(! q) Es(! k ) = k = Z 1 0 (1 v) q k df (v) e.g. k =3, q =10: 22 hence Z 1 0 (1 v) q k q 1 df (v) = Hausdor momentproblem k
Confidence bounds for randomly selected level-q subsets...and Lemma 2: For any k 2 [, ] and q 2 [k, m], P m! k =Es(!) for some! 2 q (! m ) Es(! k ) = k m k = k B(m q + k, q k + 1) q k Note:! k = { (1),..., (B(, ) =beta function) (k) } is statistically equivalent to a randomly selected subset of! m! k =Es(!) for some! 2 q(! m) Es(! k ) = k occurs i : q k of the samples in! m \! k satisfy f(x (! k ), ) apple 0 and the remaining m q samples satisfy f(x (! k ), ) > 0 hence Lem. 2 follows from Lem. 1 and the law of total probability 23
Confidence bounds for randomly selected level-q subsets...and Lemma 2: For any k 2 [, ] and q 2 [k, m], P m! k =Es(!) for some! 2 q (! m ) Es(! k ) = k m k = k B(m q + k, q k + 1) q k Note:! k = { (1),..., (B(, ) =beta function) (k) } is statistically equivalent to a randomly selected subset of! m! k =Es(!) for some! 2 q(! m) Es(! k ) = k occurs i : q k of the samples in! m \! k satisfy f(x (! k ), ) apple 0 and the remaining m q samples satisfy f(x (! k ), ) > 0 hence Lem. 2 follows from Lem. 1 and the law of total probability 23
Confidence bounds for randomly selected level-q subsets Proof of Thm. 1 (sketch): I P m! k =Es(!) for some! 2 q(! m) \ V x (! k ) apple Es(! k ) = k! = k m k B( ; m q + k, q k +1) q k (B(,, ) =incomplete beta function) I P m V x (! k ) apple! k =Es(!) for some! 2 q(! m) = Pm k! k =Es(!) for! 2 q(! m) \ V x (! k ) apple Es(! k ) = k P m k! k =Es(!) for some! 2 q(! m) Es(! k ) = k = (q k; m, 1 ) (Baye s law) I Hence P m V x (!) apple! 2 q(! m) \ Es(!) = k = (q k; m, 1 ) Thm. 1 follows from total law of probability, since, for all k 2 [, ], 24 (q ; m, 1 ) apple (q k; m, 1 ) apple (q ; m, 1 )
Confidence bounds for randomly selected level-q subsets Proof of Thm. 1 (sketch): I P m! k =Es(!) for some! 2 q(! m) \ V x (! k ) apple Es(! k ) = k! = k m k B( ; m q + k, q k +1) q k (B(,, ) =incomplete beta function) I P m V x (! k ) apple! k =Es(!) for some! 2 q(! m) = Pm k! k =Es(!) for! 2 q(! m) \ V x (! k ) apple Es(! k ) = k P m k! k =Es(!) for some! 2 q(! m) Es(! k ) = k = (q k; m, 1 ) (Baye s law) I Hence P m V x (!) apple! 2 q(! m) \ Es(!) = k = (q k; m, 1 ) Thm. 1 follows from total law of probability, since, for all k 2 [, ], 24 (q ; m, 1 ) apple (q k; m, 1 ) apple (q ; m, 1 )
Confidence bounds for randomly selected level-q subsets Proof of Thm. 1 (sketch): I P m! k =Es(!) for some! 2 q(! m) \ V x (! k ) apple Es(! k ) = k! = k m k B( ; m q + k, q k +1) q k (B(,, ) =incomplete beta function) I P m V x (! k ) apple! k =Es(!) for some! 2 q(! m) = Pm k! k =Es(!) for! 2 q(! m) \ V x (! k ) apple Es(! k ) = k P m k! k =Es(!) for some! 2 q(! m) Es(! k ) = k = (q k; m, 1 ) (Baye s law) I Hence P m V x (!) apple! 2 q(! m) \ Es(!) = k = (q k; m, 1 ) Thm. 1 follows from total law of probability, since, for all k 2 [, ], 24 (q ; m, 1 ) apple (q k; m, 1 ) apple (q ; m, 1 )
Confidence bounds for randomly selected level-q subsets Consider the relationship between optimal costs: J o ( ) of CCP J (!) of SP(!) Prop. 1: P m J (!) J o ( )! 2 q (! m ) (q ; m, 1 ) Proof: J (!) J o ( ) if the solution x of SP(!) is feasible for CCP, so P m V x (!) apple! 2 q(! m) apple P m J (! r) J o ( )! 2 q(! m) Prop. 2: P m J (!) >J o ( )! = q apple (q 1; q, 1 ) Proof: J (!) apple J o ( ) if the solution x o of CCP is feasible for SP(!) i.e. if f x o ( ), apple 0 for all 2!, so P m J (!) apple J o ( ) P f x o ( ), apple 0 (1 ) q =1 (q 1; q, 1 ) q 25
Confidence bounds for randomly selected level-q subsets Consider the relationship between optimal costs: J o ( ) of CCP J (!) of SP(!) Prop. 1: P m J (!) J o ( )! 2 q (! m ) (q ; m, 1 ) Proof: J (!) J o ( ) if the solution x of SP(!) is feasible for CCP, so P m V x (!) apple! 2 q(! m) apple P m J (! r) J o ( )! 2 q(! m) Prop. 2: P m J (!) >J o ( )! = q apple (q 1; q, 1 ) Proof: J (!) apple J o ( ) if the solution x o of CCP is feasible for SP(!) i.e. if f x o ( ), apple 0 for all 2!, so P m J (!) apple J o ( ) P f x o ( ), apple 0 (1 ) q =1 (q 1; q, 1 ) q 25
Confidence bounds for randomly selected level-q subsets Consider the relationship between optimal costs: J o ( ) of CCP J (!) of SP(!) Prop. 1: P m J (!) J o ( )! 2 q (! m ) (q ; m, 1 ) Proof: J (!) J o ( ) if the solution x of SP(!) is feasible for CCP, so P m V x (!) apple! 2 q(! m) apple P m J (! r) J o ( )! 2 q(! m) Prop. 2: P m J (!) >J o ( )! = q apple (q 1; q, 1 ) Proof: J (!) apple J o ( ) if the solution x o of CCP is feasible for SP(!) i.e. if f x o ( ), apple 0 for all 2!, so P m J (!) apple J o ( ) P f x o ( ), apple 0 (1 ) q =1 (q 1; q, 1 ) q 25
Comparison with deterministic discarding strategies For optimal sample discarding (i.e. SP) or a greedy heuristic we have P m V x (! ) apple (q, ; m, ), (q, ; m q + 1 m, ) =1 m q (m q + 1; m, ) Calafiore (2010), Campi & Garatti (2011) this bound is looser than the lower bound for random sample discarding (i.e. SP(!),! 2 q (! m )) since 1 (q, ; m q + 1 m, ) = 1 (q ; m, 1 ) m q implies B (q, ; m, ) apple (q ; m, 1 ) 26 B (q, ; m, ) = (q ; m, 1 ) only if =1or q = m (in which case q(! m) =1)
Comparison with deterministic discarding strategies For optimal sample discarding (i.e. SP) or a greedy heuristic we have P m V x (! ) apple (q, ; m, ), (q, ; m q + 1 m, ) =1 m q (m q + 1; m, ) Calafiore (2010), Campi & Garatti (2011) this bound is looser than the lower bound for random sample discarding (i.e. SP(!),! 2 q (! m )) since 1 (q, ; m q + 1 m, ) = 1 (q ; m, 1 ) m q implies B (q, ; m, ) apple (q ; m, 1 ) 26 B (q, ; m, ) = (q ; m, 1 ) only if =1or q = m (in which case q(! m) =1)
Comparison with deterministic discarding strategies The lower confidence bound for optimal sample discarding corresponds to the worst case assumption: since P m V x (!) > =0for all! 2 q (! m ) such that! 6=!? Thm. 1 implies ( 1 E m q(! m) E m X P m V x (!) >!2 q(! m) ) apple 1 (q ; m, 1 )? Lem. 2 implies E m m q + 1 q(! m) apple m q 27
Comparison with deterministic discarding strategies Confidence bounds for m = 500 with q = d0.75me = 375, =1, = 10 1 0.9 0.8 0.7 Confidence 0.6 0.5 0.4 0.3 0.2 Φ(q ζ; m, 1 ϵ) Φ(q ζ; m, 1 ϵ) 0.1 Ψ(q, ζ; m, ϵ) 0 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Violation probability ϵ 28
Comparison with deterministic discarding strategies Values of lying on 5% and 95% confidence bounds for varying m with q = d0.75me, =1, = 10 Violation probability 0.6 ϵ 95 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2 ϵ 5 ϵ 95 0.15 10 2 10 4 10 6 10 8 Sample size m 10 0 ϵ 95 ϵ 5 ϵ 95 ϵ 5 10-1 10-2 10-3 10-4 probablities 5, 95 and 0 95 are defined by 10 2 10 4 10 6 10 8 Sample size m 0.05 = (q ; m, 1 5 ), 0.95 = (q ; m, 1 95) 0.95 = (q, ; m, 0 95) 29
Example: minimal bounding circle Empirical distribution of violation probability for smallest bounding circle with: N(0,I), q = 80, m = 100, and500 realisations of! m 1 0.9 0.8 x(ω), ω Ω q (ω m ) x = x (ω ) Probability of V (x) ϵ 0.7 0.6 0.5 0.4 0.3 (78; 100, 1 ) (77; 100, 1 ) - - 0.2 0.1 (80, 3; 100, ) 0 0 0.1 0.2 0.3 0.4 0.5 Violation probability ϵ 30
Example: minimal bounding circle Empirical cost distributions for smallest bounding circle with: N(0,I), q = 80, m = 100, and500 realisations of! m 1 0.9 ω Ω q (ω m ) ω = ω Probability of J (ω) J o (ϵ) 0.8 0.7 0.6 0.5 0.4 0.3 0.2 @I (99; 100, 1 ) (77; 100, 1 ) - 0.1 (80, 3; 100, ) 0 0 0.1 0.2 0.3 0.4 0.5 Violation probability ϵ 31
Outline 1. Problem definition and motivation 2. Confidence bounds for randomised sample discarding 3. Posterior confidence bounds using empirical violation estimates 4. Randomized algorithm with prior and posterior confidence bounds 5. Conclusions and future directions 32
Random constraint selection strategy How can we select! 2 q (! m ) at random? Summary of approach: For given! m,let! r = { (1),..., (r) },and solve SP(! r ) for x (! r ) count q = 2! m : f x (! r ), apple 0 then 2! m : f x (! r ), apple 0 2 q (! m ) Choose r to maximize the probability that q lies in the desired range e.g. r =10 m =50 q =13+10=23 33
Random constraint selection strategy How can we select! 2 q (! m ) at random? Summary of approach: For given! m,let! r = { (1),..., (r) },and solve SP(! r ) for x (! r ) count q = 2! m : f x (! r ), apple 0 then 2! m : f x (! r ), apple 0 2 q (! m ) Choose r to maximize the probability that q lies in the desired range e.g. r =10 m =50 q =13+10=23 33
Random constraint selection strategy How can we select! 2 q (! m ) at random? Summary of approach: For given! m,let! r = { (1),..., (r) },and solve SP(! r ) for x (! r ) count q = 2! m : f x (! r ), apple 0 then 2! m : f x (! r ), apple 0 2 q (! m ) Choose r to maximize the probability that q lies in the desired range e.g. r =10 m =50 q =13+10=23 33
Random constraint selection strategy How can we select! 2 q (! m ) at random? Summary of approach: For given! m,let! r = { (1),..., (r) },and solve SP(! r ) for x (! r ) count q = 2! m : f x (! r ), apple 0 then 2! m : f x (! r ), apple 0 2 q (! m ) Choose r to maximize the probability that q lies in the desired range e.g. r =10 m =50 q =13+10=23 33
Posterior confidence bounds Define r (! m )= 2! m : f x (! r ), apple 0 Theorem 2: For any 2 [0, 1] and apple r apple q apple m, P m V x (! r ) apple r (! m )=q (q ; m, 1 ). For any 2 [0, 1] and apple r apple q apple m], P m V x (! r ) apple r (! m )=q apple (q ; m, 1 ).? r(! m) provides an empirical estimate of violation probability? Thm. 2 gives aposterioriconfidence bounds (i.e. for given r(! m))? These bounds are tight for general probability distributions, i.e. they cannot be improved if can take any value in [, ] they are exact for problems with = 34
Posterior confidence bounds Define r (! m )= 2! m : f x (! r ), apple 0 Theorem 2: For any 2 [0, 1] and apple r apple q apple m, P m V x (! r ) apple r (! m )=q (q ; m, 1 ). For any 2 [0, 1] and apple r apple q apple m], P m V x (! r ) apple r (! m )=q apple (q ; m, 1 ).? r(! m) provides an empirical estimate of violation probability? Thm. 2 gives aposterioriconfidence bounds (i.e. for given r(! m))? These bounds are tight for general probability distributions, i.e. they cannot be improved if can take any value in [, ] they are exact for problems with = 34
Probability of selecting a level-q subset of! m Selection of r is based on Lemma 3: For any apple r apple q apple m, m r P m r (! m )=q min q r m r P m r (! m )=q apple q r 2[, ] max 2[, ] B(m q +,q + 1) B(,r + 1) B(m q +,q + 1) B(,r + 1) Proof (sketch): 1 From Lem. 1 & Lem. 2, for k 2 [, ] we have P m V x (! r) = v Es(! r) = k = (1 v)r k v k 1 B(k, r k +1) dv 2 r(! m)=q i q r of the m r samples in! m \! r satisfy f(x (! r), ) apple 0 and the remaining m q samples satisfy f(x (! r), ) > 0 35 Hence Lem. 3 follows from 1 and the law of total probability
Probability of selecting a level-q subset of! m Selection of r is based on Lemma 3: For any apple r apple q apple m, m r P m r (! m )=q min q r m r P m r (! m )=q apple q r 2[, ] max 2[, ] B(m q +,q + 1) B(,r + 1) B(m q +,q + 1) B(,r + 1) Proof (sketch): 1 From Lem. 1 & Lem. 2, for k 2 [, ] we have P m V x (! r) = v Es(! r) = k = (1 v)r k v k 1 B(k, r k +1) dv 2 r(! m)=q i q r of the m r samples in! m \! r satisfy f(x (! r), ) apple 0 and the remaining m q samples satisfy f(x (! r), ) > 0 35 Hence Lem. 3 follows from 1 and the law of total probability
Posterior confidence bounds (contd.) Proof of Thm. 2 (sketch): I P m r(! m)=q \ V x (! r) apple Es(! r) = k! = m r B( ; m q + k, q k +1) q r B(k, r k +1) I Hence from Lem. 3 and Bayes law P m V x (! r) apple r(! m)=q \ Es(! r) = k = (q k; m, 1 ) so the total law of probability implies P m V x (! r) apple r(! m)=q (q ; m, 1 ) X k= P m Es(! r) = k P m V x (! r) apple r(! m)=q min{r, } X apple (q ; m, 1 ) k= P m Es(! r) = k 36
Posterior confidence bounds (contd.) Proof of Thm. 2 (sketch): I P m r(! m)=q \ V x (! r) apple Es(! r) = k! = m r B( ; m q + k, q k +1) q r B(k, r k +1) I Hence from Lem. 3 and Bayes law P m V x (! r) apple r(! m)=q \ Es(! r) = k = (q k; m, 1 ) so the total law of probability implies P m V x (! r) apple r(! m)=q (q ; m, 1 ) X k= P m Es(! r) = k P m V x (! r) apple r(! m)=q min{r, } X apple (q ; m, 1 ) k= P m Es(! r) = k 36
Posterior confidence bounds (contd.) Proof of Thm. 2 (sketch): I P m r(! m)=q \ V x (! r) apple Es(! r) = k! = m r B( ; m q + k, q k +1) q r B(k, r k +1) I Hence from Lem. 3 and Bayes law P m V x (! r) apple r(! m)=q \ Es(! r) = k = (q k; m, 1 ) so the total law of probability implies P m V x (! r) apple r(! m)=q (q ; m, 1 ) X k= P m Es(! r) = k P m V x (! r) apple r(! m)=q min{r, } X apple (q ; m, 1 ) k= P m Es(! r) = k 36
Outline 1. Problem definition and motivation 2. Confidence bounds for randomised sample discarding 3. Posterior confidence bounds using empirical violation estimates 4. Randomized algorithm with prior and posterior confidence bounds 5. Conclusions and future directions 37
Parameters and target confidence bounds Consider the design of an algorithm to generate ˆx: ˆq: an approximate CCP solution the number of constraints satisfied out of m sampled constraints Impose prior confidence bounds: V (ˆx) 2 (, ] with probability p prior Impose posterior confidence bounds, given the value of ˆq: V (ˆx) ˆq/m apple with probability p post 38
Parameters and target confidence bounds ˆq cannot be specified directly, but, for given m and q, q: B Lem. 3 gives the probability, p trial, of r (! m ) 2 [q, q], for given r B hence SP(! r ) must be solved N trial times in order to ensure that ˆq 2 [q, q] with a given prior probability B choose r = r, the maximizer of p trial (possibly subject to r apple r max ) in order to minimize N trial Selection of m and q, q: B choose m so that posterior confidence bounds are met: m(1 ) ; m, 1 /2 1 (1 + ppost) 2 m(1 ) ; m, 1 + /2 apple 1 (1 ppost). 2 B Thm. 2 allows q, q to be chosen so that prior confidence bounds hold 39
Randomized algorithm with tight confidence bounds Algorithm: (i). Given bounds,, probabilities p prior <p post,andsamplesizem, determine q =min q subject to (q ; m, 1 ) 1 2 (1 + p post) q = max q subject to (q ; m, 1 ) apple 1 2 (1 p post) qx m r r = arg max r q=q q r min 2[, ] B(m q +,q + 1) B(,r + 1) qx m r B(m q +,q + 1) p trial = q r min 2[, ] B(,r + 1) q=q ln(1 pprior /p post ) N trial = ln(1 p trial ) 40
Randomized algorithm with tight confidence bounds Algorithm (cont.): (ii). Draw N trial multisamples,! (i) m,andforeachi =1,...,N trial : compute x (! (i) r ) count r (! (i) m )= 2! (i) m (iii). Determine i 2{1,...,N trial }: and return i = arg min i2{1,...,n trial } : f x (! (i) r ), apple 0 1 2 ( q + q) r (!(i) m ), the solution estimate, ˆx = x (! (i ) r ) the number of satisfied constraints, ˆq = r (! (i ) m ) 40
Randomized algorithm with tight confidence bounds Proposition 3: ˆx, ˆq satisfy the prior confidence bounds P N trial m V (ˆx) 2 (, ] p prior and the posterior confidence bounds (q ; m, 1 ) apple P N trial m V (ˆx) apple ˆq = q apple (q ; m, 1 ). The posterior bounds follow directly from Thm. 2 The choice of q, q gives P m V (ˆx) 2 (, ] ˆq = q p post 8q 2 [q, q], =) P N trial m V (ˆx) 2 (, ] p post P N trial m ˆq 2 [q, q] but the definition of i implies P N trial m ˆq 2 [q, q] 1 (1 p trial ) N trial =) P N trial m V (ˆx) 2 (, ] p post 1 (1 p trial ) N trial so the choice of N trial ensures the prior bound 41
Randomized algorithm with tight confidence bounds B Large m is computationally cheap (hence small and p post 1) B p trial is a tight bound on the probability of ˆq 2 [q, q] and is exact if = B The posterior confidence bounds are tight for all distributions and are exact if = B Repeated trials and empirical violation tests have been used before e.g. Calafiore (2017), Chamanbaz et al. (2016) B The lower posterior confidence bound derived here coincides with Calafiore (2017) for fully supported problems ( = = n) but our approach: shows that this bound is exact for fully supported problems provides tight bounds in all other cases 42
Randomized algorithm with tight confidence bounds Variation of r and N trial with, for m = 10 5, =0.19, =0.21, =0.005, andp post =0.5(1 + p prior ): (, ) (2, 5) (7, 10) (17, 20) (47, 50) (97, 100) p prior p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 r 15 40 91 241 492 N trial 84 109 176 291 37 48 77 128 22 29 46 76 13 16 26 43 8111729 (, ) (1, 2) (1, 5) (1, 10) p prior p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 p 1 p 2 p 3 p 4 r 5 12 22 N trial 96 125 200 331 189 246 396 655 1022 1329 2116 3465 where p 1 =0.9, p 2 =0.95, p 3 =0.99, p 4 =0.999 43
Example I: minimal bounding hypersphere Find smallest hypersphere in R 4 containing N(0,I) with probability 0.8 Support dimension: =2, =5 Prior bounds: V (ˆx) 2 (0.19, 0.21] with probability p prior =0.9 Posterior bounds: V (ˆx) ˆq/m apple0.005 with probability p post =0.95 Parameters: m = 10 5, q = 79257, q = 80759 r = 15, p trial =0.0365, N trial = 84 Compare SP solved approximately using greedy discarding (assuming one constraint discarded per iteration), with m = 420, q = 336: V x (! ) 2 (0.173, 0.332] with probability 0.9 44
Example I: minimal bounding hypersphere Find smallest hypersphere in R 4 containing N(0,I) with probability 0.8 Support dimension: =2, =5 Prior bounds: V (ˆx) 2 (0.19, 0.21] with probability p prior =0.9 Posterior bounds: V (ˆx) ˆq/m apple0.005 with probability p post =0.95 Parameters: m = 10 5, q = 79257, q = 80759 r = 15, p trial =0.0365, N trial = 84 Compare SP solved approximately using greedy discarding (assuming one constraint discarded per iteration), with m = 420, q = 336: V x (! ) 2 (0.173, 0.332] with probability 0.9 44
Example I: minimal bounding hypersphere Posterior confidence bounds given ˆq 2 [q, q]: P{V (ˆx) apple ˆq = q} (blue) P{V (ˆx) apple ˆq = q} (red) 1 0.9 Φ( q ζ; m, 1 ϵ) Φ(q ζ; m, 1 ϵ) 0.8 0.7 Probability 0.6 0.5 0.4 0.3 0.2 0.1 0 0.185 0.19 0.195 0.2 0.205 0.21 0.215 Violation probability ϵ 45
Example I: minimal bounding hypersphere 1 0.9 Observed frequency Worst-case bound 0.8 0.7 Probability 0.6 0.5 0.4 0.3 0.2 0.1 0 0.785 0.79 0.795 0.8 0.805 0.81 0.815 Proportion of sampled constraints satisfied ˆq/m Empirical probability distribution of ˆq (blue) for 1000 repetitions Worst case bound F (red), P{ ˆq/m 0.8 applet} F (0.8+t) F (0.8 t) 46
Example II: finite horizon robust control design Control problem (from Calafiore 2017): Determine a control sequence to drive the system state to a target over a finite horizon with high probability despite model uncertainty Discrete-time uncertain linear system z(t + 1) = A( )z(t)+bu(t), A( )=A 0 + i,j 2 [ Xn z i,j=1 i,je i e > j, ] uniformly distributed 47
Example II: finite horizon robust control design Control problem (from Calafiore 2017): Determine a control sequence to drive the system state to a target over a finite horizon with high probability despite model uncertainty Discrete-time uncertain linear system z(t + 1) = A( )z(t)+bu(t), A( )=A 0 + i,j 2 [ Xn z i,j=1 i,je i e > j, ] uniformly distributed 47
Example II: finite horizon robust control design Control objective: CCP : minimize,u subject to P kz T z(t, )k 2 + kuk 2 > apple with u = {u(0),...,u(t 1)}: predicted control sequence z(t, ): = A( ) T z(0) + A( ) T 1 B B u z T : target state 0.4 0.35 0.3 control input u(t) 0.25 0.2 0.15 0.1 0.05 0-0.05 48-0.1 0 2 4 6 8 10 time step t
Example II: finite horizon robust control design Problem dimensions: n z =6, T = 10, n = 11, =1, =3 and parameters: = 10 3, =0.005 Single-sided prior probability bounds with =0.005: P V (x) 2 (0, ] p prior =0.99 Posterior uncertainty: =0.003 and probability: p post =1 10 12 =) m = 63 10 3 Choose number of sampled constraints in each trial to maximize p trial : r = 2922, p trial =0.482, N trial =9 if r is limited to 2000, we get p trial =0.414, N trial =9 49
Example II: finite horizon robust control design Compare number of trials needed for ˆq expected using bounds of Calfiore (2017): 9.6 expected using proposed approach (1/p trial ): 2.4 observed: 1.4 Residual conservativism is due to use of bounds: =1, =3 rather than the actual distribution for : support dimension: 1 2 3 observed frequency: 5% 59% 36% q 50
Future directions B Alternative methods for randomly selecting level-q solutions B Methods of checking feasibility of suboptimal points B Estimation of violation probability over a set of parameters application to computing invariant sets for constrained dynamics v2 1.5 1 0.5 0 0.5 1 1.5 1.5 1 0.5 0 0.5 1 1.5 v1 B Application to chance-constrained energy management in PHEVs 51
Conclusions B Solutions of sampled convex optimization problems with randomised sample selection strategies have tighter confidence bounds than optimal sample discarding strategies B Randomised sample selection can be performed by solving a robust sampled optimization and counting violations of additional sampled constraints B We propose a randomised algorithm for chance-constrained convex programming with tight a priori and a posteriori confidence bounds 52
Thank you for your attention! Acknowledgements: James Fleming, Matthias Lorenzen, Johannes Buerger, Rainer Manuel Schaich... 53
Research directions B Optimal control of uncertain systems chance-constrained optimization stochastic predictive control B Robust control design invariant sets and uncertainty parameterization B Applications: energy management in hybrid electric vehicles (BMW) ecosystem-based fisheries management (BAS) 54