Yishay Mansour. AT&T Labs and Tel-Aviv University. design special-purpose planning algorithms that exploit. this structure.

Size: px
Start display at page:

Download "Yishay Mansour. AT&T Labs and Tel-Aviv University. design special-purpose planning algorithms that exploit. this structure."

Transcription

1 A Sparse Sampling Algoritm for Near-Optimal Planning in Large Markov Decision Processes Micael Kearns AT&T Labs Yisay Mansour AT&T Labs and Tel-Aviv University Andrew Y. Ng UC Berkeley Abstract An issue tat is critical for te application of Markov decision processes (MDPs) to realistic problems is ow te complexity of planning scales wit te size of te MDP. In stocastic environments wit very large or even innite state spaces, traditional planning and reinforcement learning algoritms are often inapplicable, since teir running time typically scales linearly wit te state space size in te worst case. In tis paper we present a new algoritm tat, given only a generative model (simulator) for an arbitrary MDP, performs near-optimal planning wit a running time tat as no dependence on te number of states. Altoug te running time is exponential in te orizon time (wic depends only on te discount factor and te desired degree of approximation to te optimal policy), our results establis for te rst time tat tere are no teoretical barriers to computing near-optimal policies in arbitrarily large, unstructured MDPs. 1 Introduction In te past decade, Markov decision processes (MDPs) and reinforcement learning ave become a standard framework for planning and learning under uncertainty witin te articial intelligence literature. Te desire to attack problems of increasing complexity wit tis formalism as recently led researcers to focus particular attention on te case of (exponentially or even in- nitely) large state spaces. A number of interesting algoritmic and representational suggestions ave been made for coping wit suc large MDPs. Function approximation [SB98] is a well-studied approac to learning value functions in large state spaces, and many autors ave recently begun to study te properties of large MDPs tat enjoy compact representations, suc as MDPs in wic te state transition probabilities factor into a small number of components [MHK + 98]. In tis paper, we are interested in te problem of computing a near-optimal policy in a large or innite MDP tat is given tat is, we are interested in planning. It sould be clear tat as an MDP becomes very large, te classical planning assumption tat te MDP is given explicitly by tables of rewards and transition probabilities becomes infeasible. One approac to tis diculty is to assume tat te MDP as some special structure tat permits compact representation (suc as te factored transition probabilities mentioned above), and to design special-purpose planning algoritms tat exploit tis structure. Here we take a rater dierent approac. We consider a setting in wic our planning algoritm is given access to a generative model, orsimulator, of te MDP. Informally, tis is a \black box" to wic we can give any state-action pair (s; a), and receive in return a randomly sampled next state and reward from te distributions associated wit (s; a). Generative models are a natural way in wic a large MDP migt be specied, and are more general tan most structured representations, in te sense tat structured representations usually provide an ecient way of implementing a generative model. Note also tat since a generative model provides less information tan explicit tables of probabilities, but more information tan a single continuous trajectory of experience generated according to some exploration policy, results obtained via a generative model blur te distinction between wat is typically called \planning" and \learning" in MDPs. Our main result is a new algoritm tat accesses te given generative model to perform near-optimal planning in an \on-line" fasion. From any given state s, te algoritm samples te generative model for many dierent state-action pairs, and uses tese samples to compute a near-optimal action from s. Te amount of time required to compute a near-optimal action from any particular state s as no dependence on te number of states in te MDP, even toug te next-state distributions from s may of course be spread over te entire state space. Te key to our analysis is in sowing tat appropriate sparse sampling suces to construct enoug information about te environment near s to compute a near-optimal action. Te analysis relies on a combination of Bellman equation calculations, wic are standard in reinforcement learning, and uniform convergence arguments, wic are standard in supervised learning; tis combina-

2 tion of tecniques was rst applied in [KS99]. As mentioned, te running time required at eac state does ave an exponential dependence on te orizon time (wic can be sown to be unavoidable witout furter assumptions). Note tat tis learning algoritm is itself simply a (stocastic) policy tat appens to use a generative model as a subroutine. In tis sense, if we view te generative model as providing a \compact" representation of te MDP, our algoritm provides a correspondingly \compact" representation of a near-optimal policy. We view our result as complimentary to work tat proposes and exploits particular compact representations of MDPs [MHK + 98], wit bot lines of work beginning to demonstrate te potential feasibility of planning and learning in very large environments. Preliminaries We begin wit te denition of a Markov decision process on a set of N = jsj states, explicitly allowing te possibility of te number of states being (countably or uncountably) innite. Denition 1 A Markov decision process M on a set of states S and wit actions fa 1 ;:::;a k g consists of: Transition Probabilities: For eac state-action pair (s; a), a next-state distribution P sa (s 0 ) tat species te probability of transition to eac state s 0 upon execution of action a from state s. Reward Distributions: For eac state-action pair (s; a), a distribution R sa on real-valued rewards for executing action a from state s. We assume rewards are bounded in absolute value by R max. For simplicity, we sall assume in tis paper tat all rewards are in fact deterministic. However, all of our results ave easy generalizations for te case of stocastic rewards, wit an appropriate and necessary dependence on te variance of te reward distributions. Denition A generative model for a Markov decision process M is a randomized algoritm tat, on input of a state-action pair (s; a), outputs R sa and a state s 0, were s 0 is randomly drawn according to te transition probabilities P sa (). Following standard terminology in reinforcement learning, we dene a (stocastic) policy to be any mapping : S 7! fa 1 ;:::;a k g.tus (s) may be a random variable, but depends only on te current state s. We will be primarily concerned wit discounted reinforcement learning 1,sowe assume we are givenanumber 0 <1called te discount factor, wit wic we ten dene te value function V for any policy : V (s) =E " X 1 # i,1 r i s; (1) 1 However, most of our results ave straigtforward generalizations to te undiscounted nite-orizon case for any xed orizon H. were r i is te reward received on te it step of executing te policy from state s, and te expectation is over te transition probabilities and any randomization in. Note tat for any s and any, jv (s)j V max, were we dene V max = R max =(1, ). We also dene te Q-function for a given policy as Q (s; a) =R sa + E s 0P sa() [V (s 0 )] () (were te notation s 0 P sa () means tat s 0 is drawn according to te distribution P sa ()). We will later describe an algoritm A tat takes as input any state s and (stocastically) outputs an action a, and wic terefore implements a policy. Wen we ave suc anal- goritm, we will also write V A and Q A to denote te value function and Q-function of te policy implemented by A. Finally, we dene te optimal value function and te optimal Q-function as V (s) = sup V (s) and Q (s; a) = sup Q (s; a), and te optimal policy, (s) = arg max a Q (s; a) for all s S. 3 Planning in Large or Innite MDPs Usually one considers te planning problem in MDPs to be tat of computing a near-optimal policy, given as input te transition probabilities P sa () and te rewards R sa (for instance, by solving te MDP for te optimal policy). Tus, te input is a complete and exact model, and te output is a total mapping from states to actions. Witout additional assumptions about te structure of te MDP, suc an approac is clearly infeasible in very large state spaces, were even reading all of te input can take N time, and even specifying a general policy requires space on te order of N. In suc MDPs, a more fruitful way of tinking about planning migt be an online view, in wic we examine te per-state complexity of planning. Tus, te input to a planning algoritm would be a single state, and te output would be wic single action to take from tat state. In tis on-line view, a planning algoritm is itself simply a policy (but one tat may need to perform some nontrivial computation at eac state). Our main result is te description and analysis of an algoritm A tat, given access to a generative model for an arbitrary MDP M, takes any state of M as input and produces an action as output, and meets te following performance criteria: Te policy implemented by A is near-optimal in M; Te running time of A (tat is, te time required to compute an action at any state) as no dependence on te number of states of M. Tis result is obtained under te assumption tat te input state to A requires only O(1) space, a standard assumption known as te uniform cost model [AHU74], tat is typically adopted to allow analysis of algoritms tat operate on real numbers (suc as we require to allow innite state spaces). If one is unappy wit tis model, ten algoritm A will suer a dependence on te number of states only equal to te space required to name te states (at worst log(n) for N states).

3 3.1 A Sparse Sampling Planner Here is our main result: Teorem 1 Tere is a randomized algoritm A tat, given access to a generative model for any MDP M, takes as input any state s S and any value ">0, outputs an action, and satises te following two conditions: (Eciency) Te running time of A is O((kC) H ), were C = V max H = log (=V max ) ; H log khv max + log R max =((1, ) )=4; V max = R max =(1, ): ; In particular, te running time depends only on R max,, and ", and does not depend on N = jsj. If we view R max as a constant, tis can also be written k "(1, ) O, 1 1, log, 1 "(1,) : (3) (Near-Optimality) Te value function of te stocastic policy implemented by A satises jv A (s), V (s)j " (4) simultaneously for all states s S. As we ave already suggested, it will be elpful to tink of algoritm A in two dierent ways. On te one and, A is an algoritm tat takes a state as input and as access to a generative model, and as suc we sall be interested in its resource complexity its running time, and te number of calls it needs to make to te generative model (bot per state input). On te oter and, A produces an action as output in response to eac state given as input, and tus implements a (possibly stocastic) policy. Wile a sketc of te proof of Teorem 1 is given in Appendix A, and detailed pseudo-code for te algoritm is provided in Figure 1, we now give some ig-level intuition for te algoritm and its analysis. For te sake of simplicity, let us consider only te twoaction case ere, wit actions a 1 and a. Recall tat te optimal policy at s is given by (s) = arg max a Q (s; a), and terefore is completely determined by, and easily calculated from, Q (s; ). Estimating te Q-values is a common way of planning in MDPs, and te basic idea of our algoritm is to nd good estimates of Q (s; a) for all actions a by looking only witin a small neigborood of s. In particular, for our algoritm to run in time tat does not depend on N = jsj, it is critical tat te size of tis neigborood does not depend on N, even toug, for example, s may ave very diuse transition probabilities, so tat it is possible to reac any oter state in S from s. From te standard duality between Q-functions and value functions, te task of estimating Q-functions is very similar to tat of estimating value functions. So wile te algoritm uses te Q-function, we will, purely for expository purposes, actually describe ere ow we estimate V (s). Tere are two parts to te approximation we use. First, rater tan estimating V,we will actually estimate, for a value of H to be specied later, te H-step expected discounted reward V (s) =E " X # i,1 r i s; (6) were r i is te reward received on te it time step upon executing te optimal policy from s. Note te \0- step" expected discounted reward is easy to estimate: Since V 0 (s) =0,wemaysimply pick our 0-step estimates to be ^V 0 (s) = 0. Moreover, we see tat te V (s), for 1, are recursively given by V (s) = R sa + E s 0P sa ()[V,1(s 0 )] max a fr sa + E s 0Psa()[V,1 (s0 )]g (7) were a is te action taken by te optimal policy from state s. Te quality of te approximation in Equation (7) becomes better for larger values of, and is controllably tigt for te largest value = H we eventually coose. One of te main eorts in te proof is establising tat te error incurred by te recursive application of tis approximation can be made controllably small by coosing H suciently large. Tus, if we are able to obtain an estimate ^V,1 (s 0 )of V,1 (s0 ) for any s 0,we can inductively dene an algoritm for nding an estimate ^V (s) ofv (s) by making use of Equation (7). Our algoritm will approximate te expectation in Equation (7) by a sample of C random next states from te generative model, were C is a parameter to be determined (and wic, for reasons tat will become clear later, we call te \widt"). Recursively, givenaway of nding te estimator ^V,1 (s0 ) for any s 0,we nd our estimate ^V (s) ofv (s) as follows: 1. For eac action a, use te generative model to get R sa and to sample a set S a of C independently sampled states from te next-state distribution P sa ().. Use our procedure for nding ^V,1 to estimate ^V,1 (s0 ) for eac state s 0 in any of te sets S a. 3. Following Equation (7), our estimate of V (s) is ten given by ^V (s) = max a ( R sa + 1 C X s 0 Sa ^V,1(s 0 ) ) : (8) We ave described our algoritm \bottom up," but it is also informative to view it \top down." Our algoritm is essentially building a sparse look-aead tree. Figure sows a conceptual picture of tis tree for a run of te algoritm from an input state s 0, for C =3.(C will typically be muc larger.) From te root s 0,we try action a 1 tree times and action a tree times. From eac of

4 Function: EstimateQ(; C; ; G; s) Input: dept, widt C, discount, A generative model G, state s. Output: A list ( ^Q (s; a 1); ^Q (s; a );:::; ^Q (s; a k)), of estimates of te Q (s; a i). 1. If n = 0, return (0;:::;0).. For eac a A, use G to generate C samples from te next-state distribution P sa(). Let S a be a set containing tese C next-states. 3. For eac a A and let our estimate of Q (s; a) be X ^Q (s; a) 1 0 =R(s; a)+ EstimateV(, 1;C;;G;s ): (5) C s 0 S a 4. Return ( ^Q (s; a 1); ^Q (s; a );:::; ^Q (s; a k)). Function: EstimateV(; C; ; G; s) Input: dept, widt C, discount, generative model G, state s. Output: A number ^V (s) tat is an estimate of V (s). 1. Let ( ^Q (s; a 1); ^Q (s; a );:::; ^Q (s; a k)) := EstimateQ(; C; ; G; s).. Return max afa1 ;:::;a k gf ^Q (s; a)g. Function: Algoritm A(; ; R max;g;s 0) Input: tolerance, discount, max reward R max, generative model G, state s 0. Output: An action a. 1. Let te required orizon H and widt C parameters be calculated as given as functions of, and R max in Teorem1.. Let ( ^Q H(s; a 1); ^Q H(s; a );:::; ^Q H(s; a k)) := EstimateQ(H; C; ; G; s 0). 3. Return arg max afa 1;:::;a k gf ^Q H (s; a)g. Figure 1: Algoritm A for planning in large or innite state spaces. EstimateV nds te ^V EstimateQ nds analogously dened ^Q. Algoritm A implements te policy. described in te text, and te resulting states, we also try eac action C times, and so on down to dept H in te tree. Zero values assigned to te leaves ten correspond to our estimates of ^V 0, wic are \backed-up" to nd estimates of ^V 1 for teir parents, wic are in turn backed-up to teir parents, and so on, up to te root to nd an estimate of ^V H (s 0). To complete te description of te algoritm, all tat remains is to coose te dept H, dept, and C, wic controls te widt of te tree. Bounding te required dept H is te easy and standard part. It is not ard to see tat if we coose dept H = log (1, )=R max (te so-called -orizon time), ten te discounted sum of te rewards tat is obtained by considering rewards beyond tis orizon is bounded by. However, suc a tree may still be as large as M itself, depending on te coice of C. For instance, if te nextstate distribution from s is uniform or nearly uniform over all te states in M, ten it would naively seem tat, in order to approximate te next-state distributions well, wewould need to take at least C = O(N) samples, if only to make sure we see most of possible next-states at least once in our samples. Te central claim we establis about C is tat it can be cosen independent of te number of states in M, yet still result in coosing near-optimal actions at te root. Te key to te argument is tat even toug small samples may give very poor approximations to te next-state distribution at eac state in te tree, tey will, neverteless, give good estimates of te expectation terms of Equation (7), and tat is really all we need. For tis we apply a careful combination of uniform convergence metods and inductive arguments on te tree dept. Again, te tecnical details of te proof of Teorem 1 are sketced in Appendix A. Te resulting tree tus represents only a vanising fraction of all of te H-step pats starting from s 0 tat ave non-zero probability in te MDP tat is, te sparse look-aead tree covers only a vanising part of te full look-aead tree. In tis sense, our algoritm is clearly related to and inspired by classical lookaead searc tecniques [RN95] our main contribution is in sowing tat in very large stocastic environments, clever random sampling suces to reconstruct nearly all of te information available in te (exponentially or in- nitely) large full look-aead tree. Note tat in te case of deterministic environments, were from eac stateaction pair we can reac only a single next state, te sparse and full trees coincide (assuming a memoization trick described below), and our algoritm reduces to classical deterministic look-aead searc.

5 a1... a1 a a1 a a1 a a1 a a a s0 a a1 a a1 a Dept H Figure : Sparse look-aead tree of states constructed by te algoritm. (Sown wit C = 3, actions a 1, a.) 3. Practical Issues and Lower Bounds Even toug te running time of algoritm A does not depend on te size of te MDP, it still runs in time exponential in te -orizon time H, and terefore exponential in 1=(1, ). It would seem tat te algoritm would be practical only if is not too close to 1. Neverteless, tere are a couple of simple tricks tat may elp to reduce te running time in certain cases. Te rst idea is simply to use memoization in our subroutines for calculating te ^V (s)'s. In Figure, tis means tat wenever tere are two nodes at te same level of te tree tat correspond to te same state, we collapse tem into one node (keeping just one of teir subtrees). Wile it is straigtforward to sow te correctness of suc memoization procedures for deterministic procedures, one sould be careful wen addressing randomized procedures; we can sow tat te properties of te algoritm are maintained under tis optimization (details are deferred to te full version of te paper). In implementing te algoritm, one may also wis not to specify in advance, but rater just try to do as well as is possible wit te computational resources available, in wic case an \iterative-deepening" approac maybe taken. In our case, tis would entail simultaneously increasing C and H by decreasing te target. Also, as studied in Davies et. al. [DNM98], ifweave access to an initial estimate of te value function, we can replace our estimates ^V 0 (s) = 0 at te leaves wit te estimated value function at tose states. Toug we sall not do so ere, it is again easy to make formal performance guarantees depending on C, H and te supremum error of te value function estimate we are using. Unfortunately, despite tese tricks, it is not dicult to prove alower bound tat sows tat any planning algoritm wit access only to a generative model, and wic implements a policy tat is -close to optimal in a general MDP, must ave running time at least exponential in te -orizon time. 4 Summary and Related Work We ave described an algoritm for near-optimal planning from a generative model, tat as a per-state running time tat does not depend on te size of te state space, but wic is still exponential in te -orizon time. Two interesting directions for improvement are to allow partially observable MDPs, and to nd more ecient algoritms tat do not ave exponential dependence on te orizon time. As a rst step towards bot of tese goals, in a separate paper we investigate a framework in wic te goal is to use a generative model to nd a near-best strategy witin a restricted class of strategies for a POMDP. Typical examples of suc restricted strategy classes include limited-memory strategies in POMDPs, or policies in large MDPs tat implement a linear mapping from state vectors to actions. Our main result in tis framework says tat as long as te restricted class of strategies is not too \complex" (were tis is formalized using appropriate generalizations of standard notions like VC dimension from supervised learning), ten it is possible to nd a nearbest strategy from witin te class, in time tat again as no dependence on te size of te state space. If te restricted class of strategies is smootly parameterized, ten tis furter leads to a number of fast, practical algoritms for doing gradient descent to nd te nearbest strategy witin te class, were te running time of eac gradient descent step now as only linear rater tan exponential dependence on te orizon time. References [AHU74] A.V. Ao, J.E. Hopcroft, and J.D. Ullman. Te Design and Analysis of Computer Algoritms. Addison-Wesley, [DNM98] Scott Davies, Andrew Y. Ng, and Andrew Moore. Applying online-searc to reinforcement learning. In Proceedings of AAAI-98, pages 753{760. AAAI Press, [KS99] Micael Kearns and Satinder Sing. Finitesample convergence rates for Q-learning and indirect algoritms. In Neural Information Processing Systems 1. MIT Press, (to appear), [MHK + 98] N. Meuleau, M. Hauskrect, K-E. Kim, L. Peskin, L.P. Kaelbling, T. Dean, and C. Boutilier. Solving very large weakly coupled Markov decision processes. In Proceedings of AAAI, pages 165{17, [RN95] S. Russell and P. Norvig. Articial Intelligence AModern Approac. Prentice Hall, [SB98] Ricard S. Sutton and Andrew G. Barto. Reinforcement Learning. MIT Press, [SY94] Satinder Sing and Ricard Yee. An upper bound on te loss from approximate optimalvalue functions. Macine Learning, 16:7{33, Appendix A: Proof Sketc of Teorem 1 In tis appendix, we sketc te proof of Teorem 1. Trougout te analysis we will rely on te pseudo-code provided for algoritm A given in Figure 1. Te claim on te running time is immediate from te denition of algoritm A. Eac call to EstimateQ generates kc calls to EstimateV, C calls for eac action.

6 Eac recursive call also reduces te dept parameter by one, so te dept of te recursion is at most H. Terefore te running time is O((kC) H ). Te main eort is in sowing tat te values of EstimateQ are indeed good estimates of Q for te cosen values of C and H. Tere are two sources of inaccuracy in tese estimates. Te rst is tat we use only a nite sample to approximate an expectation we draw only C states from te next-state distributions. Te second source of inaccuracy is tat in computing EstimateQ, we are not actually using te values of V () but rater values returned by EstimateV, wic are temselves only estimates. Te crucial step in te proof is to sow tat as increases, te overall inaccuracy decreases. Let us rst dene an intermediate random variable tat will capture te inaccuracy due to te limited sampling. Dene U (s; a) as follows: U (s; a) =R sa + 1 C CX V (s i ) (9) were te s i are drawn according to P sa (). Note tat U (s; a) isaveraging values of V (), te unknown value function. Since U (s; a) is used only for te proof and not in te algoritm, tere is no problem in dening it tis way. Te next lemma (proof omitted) sows tat wit ig probability, te dierence between U (s; a) and Q (s; a) is at most. Lemma For any state s and action a, wit probability at least 1, e, C=V max we ave jq (s; a), U (s; a)j = E spsa()[v (s)], 1 C X i V (s i ) ; were te probability is taken over te draw of te s i from P sa (). Now tat we ave quantied te error due to nite sampling, we can bound te error from our using values returned by EstimateV rater tan V (). We bound tis error as te dierence between U (s; a) and EstimateV. In order to make our notation simpler, let V n (s) be te value returned by EstimateV(n; C; ; G; s), and let Q n (s; a) be te component in te output of EstimateQ(n; C; ; G; s) tat corresponds to action a. Using tis notation, our algoritm computes Q n (s; a) =R sa + 1 C CX V n,1 (s i ) (10) were V n,1 (s) = max a fq n,1 (s; a)g, and Q 0 (s; a) =0 for every state s and action a. We now dene a parameter n tat will eventually bound te dierence between Q (s; a) and Q n (s; a). We dene n recursively: n+1 = ( + n ) (11) were 0 = V max. Solving for H we obtain H = HX i! + H V max 1, + H V max : (1) Te next lemma (proof omitted) bounds te error in te estimation, at level n, by n. Intuitively, te error due to nite sampling contributes, wile te errors in estimation contribute n. Te combined error is + n, but since we are discounting, te eective error is only ( + n ), wic by denition is n+1. Lemma 3 Wit probability at least 1, (kc) n e, C=V max we ave tat jq (s; a), Q n (s; a)j n : (13) From H H V max + =(1, ), we also see tat for H = log (=V max ), wit probability 1, (kc) H e, C=V max all te nal estimates Q H (s 0 ;a) are witin =(1, ) from te true Q-values. Te next step is to coose C suc tat = =R max (kc) H e, C=V max will bound te probability of a bad estimate during te entire computation. Specically, C = V max H log khv max + log 1 (14) is sucient to ensure tat wit probability 1, all te estimates are accurate. At tis point weave sown tat wit ig probability, algoritm A computes a good estimate of Q (s 0 ;a) for all a, were s 0 is te input state. To complete te proof, we need to relate tis to te expected value of a stocastic policy. We give a fairly general result about MDPs, wic does not depend on our specic algoritm. (A similar result appears in [SY94].) Lemma 4 Assume tat is a stocastic policy, so tat (s) is a random variable. If for eac state s, te probability tat Q (s; (s)),q (s; (s)) <is at least 1,, ten te discounted innite orizon return of is at most (+V max )=(1,) from te optimal return, tat is, for any state sv (s), V (s) ( +V max )=(1, ). Now we can combine all te lemmas to prove our main teorem. Proof of Teorem 1: As discussed before, te running time is immediate from te algoritm, and te main work is sowing tat we compute a near-optimal policy. By Lemma 3 we ave tat te error in te estimation of Q is at most H, wit probability1,(kc) H. Using te values we cose for C and H we ave tat wit probability1, te error is at most =(1, ). By Lemma 4 tis implies tat suc a policy as te property tat from every state s, V (s), V (s) (1, ) + V max 1, : (15) Substituting back te values of = =R max and = (1, ) =4 tat we ad cosen, it follows tat V (s), V 4 (s) = : (16) (1, )

tainty within the articial intelligence literature. The desire to attack problems of increasing complexity with this formalism has recently led resear

tainty within the articial intelligence literature. The desire to attack problems of increasing complexity with this formalism has recently led resear A Sparse Sampling Algorithm for Near-Optimal Planning in Large Markov Decision Processes Michael Kearns Syntek Capital mkearns@cis.upenn.edu Yishay Mansour Tel Aviv University mansour@math.tau.ac.il October

More information

Online Learning: Bandit Setting

Online Learning: Bandit Setting Online Learning: Bandit Setting Daniel asabi Summer 04 Last Update: October 0, 06 Introduction [TODO Bandits. Stocastic setting Suppose tere exists unknown distributions ν,..., ν, suc tat te loss at eac

More information

Te comparison of dierent models M i is based on teir relative probabilities, wic can be expressed, again using Bayes' teorem, in terms of prior probab

Te comparison of dierent models M i is based on teir relative probabilities, wic can be expressed, again using Bayes' teorem, in terms of prior probab To appear in: Advances in Neural Information Processing Systems 9, eds. M. C. Mozer, M. I. Jordan and T. Petsce. MIT Press, 997 Bayesian Model Comparison by Monte Carlo Caining David Barber D.Barber@aston.ac.uk

More information

lecture 26: Richardson extrapolation

lecture 26: Richardson extrapolation 43 lecture 26: Ricardson extrapolation 35 Ricardson extrapolation, Romberg integration Trougout numerical analysis, one encounters procedures tat apply some simple approximation (eg, linear interpolation)

More information

7 Semiparametric Methods and Partially Linear Regression

7 Semiparametric Methods and Partially Linear Regression 7 Semiparametric Metods and Partially Linear Regression 7. Overview A model is called semiparametric if it is described by and were is nite-dimensional (e.g. parametric) and is in nite-dimensional (nonparametric).

More information

Copyright c 2008 Kevin Long

Copyright c 2008 Kevin Long Lecture 4 Numerical solution of initial value problems Te metods you ve learned so far ave obtained closed-form solutions to initial value problems. A closedform solution is an explicit algebriac formula

More information

Differential Calculus (The basics) Prepared by Mr. C. Hull

Differential Calculus (The basics) Prepared by Mr. C. Hull Differential Calculus Te basics) A : Limits In tis work on limits, we will deal only wit functions i.e. tose relationsips in wic an input variable ) defines a unique output variable y). Wen we work wit

More information

LIMITATIONS OF EULER S METHOD FOR NUMERICAL INTEGRATION

LIMITATIONS OF EULER S METHOD FOR NUMERICAL INTEGRATION LIMITATIONS OF EULER S METHOD FOR NUMERICAL INTEGRATION LAURA EVANS.. Introduction Not all differential equations can be explicitly solved for y. Tis can be problematic if we need to know te value of y

More information

Average Rate of Change

Average Rate of Change Te Derivative Tis can be tougt of as an attempt to draw a parallel (pysically and metaporically) between a line and a curve, applying te concept of slope to someting tat isn't actually straigt. Te slope

More information

Regularized Regression

Regularized Regression Regularized Regression David M. Blei Columbia University December 5, 205 Modern regression problems are ig dimensional, wic means tat te number of covariates p is large. In practice statisticians regularize

More information

2.8 The Derivative as a Function

2.8 The Derivative as a Function .8 Te Derivative as a Function Typically, we can find te derivative of a function f at many points of its domain: Definition. Suppose tat f is a function wic is differentiable at every point of an open

More information

1 The concept of limits (p.217 p.229, p.242 p.249, p.255 p.256) 1.1 Limits Consider the function determined by the formula 3. x since at this point

1 The concept of limits (p.217 p.229, p.242 p.249, p.255 p.256) 1.1 Limits Consider the function determined by the formula 3. x since at this point MA00 Capter 6 Calculus and Basic Linear Algebra I Limits, Continuity and Differentiability Te concept of its (p.7 p.9, p.4 p.49, p.55 p.56). Limits Consider te function determined by te formula f Note

More information

A = h w (1) Error Analysis Physics 141

A = h w (1) Error Analysis Physics 141 Introduction In all brances of pysical science and engineering one deals constantly wit numbers wic results more or less directly from experimental observations. Experimental observations always ave inaccuracies.

More information

Inf sup testing of upwind methods

Inf sup testing of upwind methods INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING Int. J. Numer. Met. Engng 000; 48:745 760 Inf sup testing of upwind metods Klaus-Jurgen Bate 1; ;, Dena Hendriana 1, Franco Brezzi and Giancarlo

More information

Poisson Equation in Sobolev Spaces

Poisson Equation in Sobolev Spaces Poisson Equation in Sobolev Spaces OcMountain Dayligt Time. 6, 011 Today we discuss te Poisson equation in Sobolev spaces. It s existence, uniqueness, and regularity. Weak Solution. u = f in, u = g on

More information

The derivative function

The derivative function Roberto s Notes on Differential Calculus Capter : Definition of derivative Section Te derivative function Wat you need to know already: f is at a point on its grap and ow to compute it. Wat te derivative

More information

Efficient algorithms for for clone items detection

Efficient algorithms for for clone items detection Efficient algoritms for for clone items detection Raoul Medina, Caroline Noyer, and Olivier Raynaud Raoul Medina, Caroline Noyer and Olivier Raynaud LIMOS - Université Blaise Pascal, Campus universitaire

More information

Introduction to Machine Learning. Recitation 8. w 2, b 2. w 1, b 1. z 0 z 1. The function we want to minimize is the loss over all examples: f =

Introduction to Machine Learning. Recitation 8. w 2, b 2. w 1, b 1. z 0 z 1. The function we want to minimize is the loss over all examples: f = Introduction to Macine Learning Lecturer: Regev Scweiger Recitation 8 Fall Semester Scribe: Regev Scweiger 8.1 Backpropagation We will develop and review te backpropagation algoritm for neural networks.

More information

Quantum Numbers and Rules

Quantum Numbers and Rules OpenStax-CNX module: m42614 1 Quantum Numbers and Rules OpenStax College Tis work is produced by OpenStax-CNX and licensed under te Creative Commons Attribution License 3.0 Abstract Dene quantum number.

More information

7.1 Using Antiderivatives to find Area

7.1 Using Antiderivatives to find Area 7.1 Using Antiderivatives to find Area Introduction finding te area under te grap of a nonnegative, continuous function f In tis section a formula is obtained for finding te area of te region bounded between

More information

arxiv: v3 [cs.ds] 4 Aug 2017

arxiv: v3 [cs.ds] 4 Aug 2017 Non-preemptive Sceduling in a Smart Grid Model and its Implications on Macine Minimization Fu-Hong Liu 1, Hsiang-Hsuan Liu 1,2, and Prudence W.H. Wong 2 1 Department of Computer Science, National Tsing

More information

Preface. Here are a couple of warnings to my students who may be here to get a copy of what happened on a day that you missed.

Preface. Here are a couple of warnings to my students who may be here to get a copy of what happened on a day that you missed. Preface Here are my online notes for my course tat I teac ere at Lamar University. Despite te fact tat tese are my class notes, tey sould be accessible to anyone wanting to learn or needing a refreser

More information

Continuity and Differentiability Worksheet

Continuity and Differentiability Worksheet Continuity and Differentiability Workseet (Be sure tat you can also do te grapical eercises from te tet- Tese were not included below! Typical problems are like problems -3, p. 6; -3, p. 7; 33-34, p. 7;

More information

Consider a function f we ll specify which assumptions we need to make about it in a minute. Let us reformulate the integral. 1 f(x) dx.

Consider a function f we ll specify which assumptions we need to make about it in a minute. Let us reformulate the integral. 1 f(x) dx. Capter 2 Integrals as sums and derivatives as differences We now switc to te simplest metods for integrating or differentiating a function from its function samples. A careful study of Taylor expansions

More information

What we learned last time

What we learned last time Wat we learned last time Value-function approximation by stocastic gradient descent enables RL to be applied to arbitrarily large state spaces Most algoritms just carry over Targets from tabular case Wit

More information

Mathematics 5 Worksheet 11 Geometry, Tangency, and the Derivative

Mathematics 5 Worksheet 11 Geometry, Tangency, and the Derivative Matematics 5 Workseet 11 Geometry, Tangency, and te Derivative Problem 1. Find te equation of a line wit slope m tat intersects te point (3, 9). Solution. Te equation for a line passing troug a point (x

More information

Polynomial Interpolation

Polynomial Interpolation Capter 4 Polynomial Interpolation In tis capter, we consider te important problem of approximating a function f(x, wose values at a set of distinct points x, x, x 2,,x n are known, by a polynomial P (x

More information

Near-Optimal conversion of Hardness into Pseudo-Randomness

Near-Optimal conversion of Hardness into Pseudo-Randomness Near-Optimal conversion of Hardness into Pseudo-Randomness Russell Impagliazzo Computer Science and Engineering UC, San Diego 9500 Gilman Drive La Jolla, CA 92093-0114 russell@cs.ucsd.edu Ronen Saltiel

More information

Cubic Functions: Local Analysis

Cubic Functions: Local Analysis Cubic function cubing coefficient Capter 13 Cubic Functions: Local Analysis Input-Output Pairs, 378 Normalized Input-Output Rule, 380 Local I-O Rule Near, 382 Local Grap Near, 384 Types of Local Graps

More information

The Complexity of Computing the MCD-Estimator

The Complexity of Computing the MCD-Estimator Te Complexity of Computing te MCD-Estimator Torsten Bernolt Lerstul Informatik 2 Universität Dortmund, Germany torstenbernolt@uni-dortmundde Paul Fiscer IMM, Danisc Tecnical University Kongens Lyngby,

More information

Order of Accuracy. ũ h u Ch p, (1)

Order of Accuracy. ũ h u Ch p, (1) Order of Accuracy 1 Terminology We consider a numerical approximation of an exact value u. Te approximation depends on a small parameter, wic can be for instance te grid size or time step in a numerical

More information

Complexity of Decoding Positive-Rate Reed-Solomon Codes

Complexity of Decoding Positive-Rate Reed-Solomon Codes Complexity of Decoding Positive-Rate Reed-Solomon Codes Qi Ceng 1 and Daqing Wan 1 Scool of Computer Science Te University of Oklaoma Norman, OK73019 Email: qceng@cs.ou.edu Department of Matematics University

More information

Math 312 Lecture Notes Modeling

Math 312 Lecture Notes Modeling Mat 3 Lecture Notes Modeling Warren Weckesser Department of Matematics Colgate University 5 7 January 006 Classifying Matematical Models An Example We consider te following scenario. During a storm, a

More information

Pre-Calculus Review Preemptive Strike

Pre-Calculus Review Preemptive Strike Pre-Calculus Review Preemptive Strike Attaced are some notes and one assignment wit tree parts. Tese are due on te day tat we start te pre-calculus review. I strongly suggest reading troug te notes torougly

More information

Polynomial Interpolation

Polynomial Interpolation Capter 4 Polynomial Interpolation In tis capter, we consider te important problem of approximatinga function fx, wose values at a set of distinct points x, x, x,, x n are known, by a polynomial P x suc

More information

MVT and Rolle s Theorem

MVT and Rolle s Theorem AP Calculus CHAPTER 4 WORKSHEET APPLICATIONS OF DIFFERENTIATION MVT and Rolle s Teorem Name Seat # Date UNLESS INDICATED, DO NOT USE YOUR CALCULATOR FOR ANY OF THESE QUESTIONS In problems 1 and, state

More information

Financial Econometrics Prof. Massimo Guidolin

Financial Econometrics Prof. Massimo Guidolin CLEFIN A.A. 2010/2011 Financial Econometrics Prof. Massimo Guidolin A Quick Review of Basic Estimation Metods 1. Were te OLS World Ends... Consider two time series 1: = { 1 2 } and 1: = { 1 2 }. At tis

More information

NUMERICAL DIFFERENTIATION. James T. Smith San Francisco State University. In calculus classes, you compute derivatives algebraically: for example,

NUMERICAL DIFFERENTIATION. James T. Smith San Francisco State University. In calculus classes, you compute derivatives algebraically: for example, NUMERICAL DIFFERENTIATION James T Smit San Francisco State University In calculus classes, you compute derivatives algebraically: for example, f( x) = x + x f ( x) = x x Tis tecnique requires your knowing

More information

2.3 Product and Quotient Rules

2.3 Product and Quotient Rules .3. PRODUCT AND QUOTIENT RULES 75.3 Product and Quotient Rules.3.1 Product rule Suppose tat f and g are two di erentiable functions. Ten ( g (x)) 0 = f 0 (x) g (x) + g 0 (x) See.3.5 on page 77 for a proof.

More information

Notes on wavefunctions II: momentum wavefunctions

Notes on wavefunctions II: momentum wavefunctions Notes on wavefunctions II: momentum wavefunctions and uncertainty Te state of a particle at any time is described by a wavefunction ψ(x). Tese wavefunction must cange wit time, since we know tat particles

More information

Exam 1 Review Solutions

Exam 1 Review Solutions Exam Review Solutions Please also review te old quizzes, and be sure tat you understand te omework problems. General notes: () Always give an algebraic reason for your answer (graps are not sufficient),

More information

Lecture 15. Interpolation II. 2 Piecewise polynomial interpolation Hermite splines

Lecture 15. Interpolation II. 2 Piecewise polynomial interpolation Hermite splines Lecture 5 Interpolation II Introduction In te previous lecture we focused primarily on polynomial interpolation of a set of n points. A difficulty we observed is tat wen n is large, our polynomial as to

More information

Sin, Cos and All That

Sin, Cos and All That Sin, Cos and All Tat James K. Peterson Department of Biological Sciences and Department of Matematical Sciences Clemson University Marc 9, 2017 Outline Sin, Cos and all tat! A New Power Rule Derivatives

More information

Differentiation in higher dimensions

Differentiation in higher dimensions Capter 2 Differentiation in iger dimensions 2.1 Te Total Derivative Recall tat if f : R R is a 1-variable function, and a R, we say tat f is differentiable at x = a if and only if te ratio f(a+) f(a) tends

More information

Some Review Problems for First Midterm Mathematics 1300, Calculus 1

Some Review Problems for First Midterm Mathematics 1300, Calculus 1 Some Review Problems for First Midterm Matematics 00, Calculus. Consider te trigonometric function f(t) wose grap is sown below. Write down a possible formula for f(t). Tis function appears to be an odd,

More information

Polynomials 3: Powers of x 0 + h

Polynomials 3: Powers of x 0 + h near small binomial Capter 17 Polynomials 3: Powers of + Wile it is easy to compute wit powers of a counting-numerator, it is a lot more difficult to compute wit powers of a decimal-numerator. EXAMPLE

More information

Chapter 5 FINITE DIFFERENCE METHOD (FDM)

Chapter 5 FINITE DIFFERENCE METHOD (FDM) MEE7 Computer Modeling Tecniques in Engineering Capter 5 FINITE DIFFERENCE METHOD (FDM) 5. Introduction to FDM Te finite difference tecniques are based upon approximations wic permit replacing differential

More information

Taylor Series and the Mean Value Theorem of Derivatives

Taylor Series and the Mean Value Theorem of Derivatives 1 - Taylor Series and te Mean Value Teorem o Derivatives Te numerical solution o engineering and scientiic problems described by matematical models oten requires solving dierential equations. Dierential

More information

Function Composition and Chain Rules

Function Composition and Chain Rules Function Composition and s James K. Peterson Department of Biological Sciences and Department of Matematical Sciences Clemson University Marc 8, 2017 Outline 1 Function Composition and Continuity 2 Function

More information

1. Consider the trigonometric function f(t) whose graph is shown below. Write down a possible formula for f(t).

1. Consider the trigonometric function f(t) whose graph is shown below. Write down a possible formula for f(t). . Consider te trigonometric function f(t) wose grap is sown below. Write down a possible formula for f(t). Tis function appears to be an odd, periodic function tat as been sifted upwards, so we will use

More information

SECTION 1.10: DIFFERENCE QUOTIENTS LEARNING OBJECTIVES

SECTION 1.10: DIFFERENCE QUOTIENTS LEARNING OBJECTIVES (Section.0: Difference Quotients).0. SECTION.0: DIFFERENCE QUOTIENTS LEARNING OBJECTIVES Define average rate of cange (and average velocity) algebraically and grapically. Be able to identify, construct,

More information

ch (for some fixed positive number c) reaching c

ch (for some fixed positive number c) reaching c GSTF Journal of Matematics Statistics and Operations Researc (JMSOR) Vol. No. September 05 DOI 0.60/s4086-05-000-z Nonlinear Piecewise-defined Difference Equations wit Reciprocal and Cubic Terms Ramadan

More information

IEOR 165 Lecture 10 Distribution Estimation

IEOR 165 Lecture 10 Distribution Estimation IEOR 165 Lecture 10 Distribution Estimation 1 Motivating Problem Consider a situation were we ave iid data x i from some unknown distribution. One problem of interest is estimating te distribution tat

More information

2.11 That s So Derivative

2.11 That s So Derivative 2.11 Tat s So Derivative Introduction to Differential Calculus Just as one defines instantaneous velocity in terms of average velocity, we now define te instantaneous rate of cange of a function at a point

More information

Flavius Guiaş. X(t + h) = X(t) + F (X(s)) ds.

Flavius Guiaş. X(t + h) = X(t) + F (X(s)) ds. Numerical solvers for large systems of ordinary differential equations based on te stocastic direct simulation metod improved by te and Runge Kutta principles Flavius Guiaş Abstract We present a numerical

More information

3.4 Worksheet: Proof of the Chain Rule NAME

3.4 Worksheet: Proof of the Chain Rule NAME Mat 1170 3.4 Workseet: Proof of te Cain Rule NAME Te Cain Rule So far we are able to differentiate all types of functions. For example: polynomials, rational, root, and trigonometric functions. We are

More information

Teaching Differentiation: A Rare Case for the Problem of the Slope of the Tangent Line

Teaching Differentiation: A Rare Case for the Problem of the Slope of the Tangent Line Teacing Differentiation: A Rare Case for te Problem of te Slope of te Tangent Line arxiv:1805.00343v1 [mat.ho] 29 Apr 2018 Roman Kvasov Department of Matematics University of Puerto Rico at Aguadilla Aguadilla,

More information

Recall from our discussion of continuity in lecture a function is continuous at a point x = a if and only if

Recall from our discussion of continuity in lecture a function is continuous at a point x = a if and only if Computational Aspects of its. Keeping te simple simple. Recall by elementary functions we mean :Polynomials (including linear and quadratic equations) Eponentials Logaritms Trig Functions Rational Functions

More information

REVIEW LAB ANSWER KEY

REVIEW LAB ANSWER KEY REVIEW LAB ANSWER KEY. Witout using SN, find te derivative of eac of te following (you do not need to simplify your answers): a. f x 3x 3 5x x 6 f x 3 3x 5 x 0 b. g x 4 x x x notice te trick ere! x x g

More information

Time (hours) Morphine sulfate (mg)

Time (hours) Morphine sulfate (mg) Mat Xa Fall 2002 Review Notes Limits and Definition of Derivative Important Information: 1 According to te most recent information from te Registrar, te Xa final exam will be eld from 9:15 am to 12:15

More information

A Reconsideration of Matter Waves

A Reconsideration of Matter Waves A Reconsideration of Matter Waves by Roger Ellman Abstract Matter waves were discovered in te early 20t century from teir wavelengt, predicted by DeBroglie, Planck's constant divided by te particle's momentum,

More information

MAT244 - Ordinary Di erential Equations - Summer 2016 Assignment 2 Due: July 20, 2016

MAT244 - Ordinary Di erential Equations - Summer 2016 Assignment 2 Due: July 20, 2016 MAT244 - Ordinary Di erential Equations - Summer 206 Assignment 2 Due: July 20, 206 Full Name: Student #: Last First Indicate wic Tutorial Section you attend by filling in te appropriate circle: Tut 0

More information

2.1 THE DEFINITION OF DERIVATIVE

2.1 THE DEFINITION OF DERIVATIVE 2.1 Te Derivative Contemporary Calculus 2.1 THE DEFINITION OF DERIVATIVE 1 Te grapical idea of a slope of a tangent line is very useful, but for some uses we need a more algebraic definition of te derivative

More information

How to Find the Derivative of a Function: Calculus 1

How to Find the Derivative of a Function: Calculus 1 Introduction How to Find te Derivative of a Function: Calculus 1 Calculus is not an easy matematics course Te fact tat you ave enrolled in suc a difficult subject indicates tat you are interested in te

More information

Bounds on the Moments for an Ensemble of Random Decision Trees

Bounds on the Moments for an Ensemble of Random Decision Trees Noname manuscript No. (will be inserted by te editor) Bounds on te Moments for an Ensemble of Random Decision Trees Amit Durandar Received: Sep. 17, 2013 / Revised: Mar. 04, 2014 / Accepted: Jun. 30, 2014

More information

Physically Based Modeling: Principles and Practice Implicit Methods for Differential Equations

Physically Based Modeling: Principles and Practice Implicit Methods for Differential Equations Pysically Based Modeling: Principles and Practice Implicit Metods for Differential Equations David Baraff Robotics Institute Carnegie Mellon University Please note: Tis document is 997 by David Baraff

More information

Optimal parameters for a hierarchical grid data structure for contact detection in arbitrarily polydisperse particle systems

Optimal parameters for a hierarchical grid data structure for contact detection in arbitrarily polydisperse particle systems Comp. Part. Mec. 04) :357 37 DOI 0.007/s4057-04-000-9 Optimal parameters for a ierarcical grid data structure for contact detection in arbitrarily polydisperse particle systems Dinant Krijgsman Vitaliy

More information

to the data. The search procedure tries to identify network structures with high scores. Heckerman

to the data. The search procedure tries to identify network structures with high scores. Heckerman 2 Learning Bayesian Networks is NP-Complete David Maxwell Cickering Computer Science Department University of California at Los Angeles dmax@cs.ucla.edu ABSTRACT Algoritms for learning Bayesian networks

More information

HOMEWORK HELP 2 FOR MATH 151

HOMEWORK HELP 2 FOR MATH 151 HOMEWORK HELP 2 FOR MATH 151 Here we go; te second round of omework elp. If tere are oters you would like to see, let me know! 2.4, 43 and 44 At wat points are te functions f(x) and g(x) = xf(x)continuous,

More information

estimate results from a recursive sceme tat generalizes te algoritms of Efron (967), Turnbull (976) and Li et al (997) by kernel smooting te data at e

estimate results from a recursive sceme tat generalizes te algoritms of Efron (967), Turnbull (976) and Li et al (997) by kernel smooting te data at e A kernel density estimate for interval censored data Tierry Ducesne and James E Staord y Abstract In tis paper we propose a kernel density estimate for interval-censored data It retains te simplicity andintuitive

More information

0.1 Differentiation Rules

0.1 Differentiation Rules 0.1 Differentiation Rules From our previous work we ve seen tat it can be quite a task to calculate te erivative of an arbitrary function. Just working wit a secon-orer polynomial tings get pretty complicate

More information

Math 1210 Midterm 1 January 31st, 2014

Math 1210 Midterm 1 January 31st, 2014 Mat 110 Midterm 1 January 1st, 01 Tis exam consists of sections, A and B. Section A is conceptual, wereas section B is more computational. Te value of every question is indicated at te beginning of it.

More information

1 Proving the Fundamental Theorem of Statistical Learning

1 Proving the Fundamental Theorem of Statistical Learning THEORETICAL MACHINE LEARNING COS 5 LECTURE #7 APRIL 5, 6 LECTURER: ELAD HAZAN NAME: FERMI MA ANDDANIEL SUO oving te Fundaental Teore of Statistical Learning In tis section, we prove te following: Teore.

More information

EFFICIENCY OF MODEL-ASSISTED REGRESSION ESTIMATORS IN SAMPLE SURVEYS

EFFICIENCY OF MODEL-ASSISTED REGRESSION ESTIMATORS IN SAMPLE SURVEYS Statistica Sinica 24 2014, 395-414 doi:ttp://dx.doi.org/10.5705/ss.2012.064 EFFICIENCY OF MODEL-ASSISTED REGRESSION ESTIMATORS IN SAMPLE SURVEYS Jun Sao 1,2 and Seng Wang 3 1 East Cina Normal University,

More information

Investigating Euler s Method and Differential Equations to Approximate π. Lindsay Crowl August 2, 2001

Investigating Euler s Method and Differential Equations to Approximate π. Lindsay Crowl August 2, 2001 Investigating Euler s Metod and Differential Equations to Approximate π Lindsa Crowl August 2, 2001 Tis researc paper focuses on finding a more efficient and accurate wa to approximate π. Suppose tat x

More information

Long Term Time Series Prediction with Multi-Input Multi-Output Local Learning

Long Term Time Series Prediction with Multi-Input Multi-Output Local Learning Long Term Time Series Prediction wit Multi-Input Multi-Output Local Learning Gianluca Bontempi Macine Learning Group, Département d Informatique Faculté des Sciences, ULB, Université Libre de Bruxelles

More information

Lab 6 Derivatives and Mutant Bacteria

Lab 6 Derivatives and Mutant Bacteria Lab 6 Derivatives and Mutant Bacteria Date: September 27, 20 Assignment Due Date: October 4, 20 Goal: In tis lab you will furter explore te concept of a derivative using R. You will use your knowledge

More information

Technology-Independent Design of Neurocomputers: The Universal Field Computer 1

Technology-Independent Design of Neurocomputers: The Universal Field Computer 1 Tecnology-Independent Design of Neurocomputers: Te Universal Field Computer 1 Abstract Bruce J. MacLennan Computer Science Department Naval Postgraduate Scool Monterey, CA 9393 We argue tat AI is moving

More information

Math 161 (33) - Final exam

Math 161 (33) - Final exam Name: Id #: Mat 161 (33) - Final exam Fall Quarter 2015 Wednesday December 9, 2015-10:30am to 12:30am Instructions: Prob. Points Score possible 1 25 2 25 3 25 4 25 TOTAL 75 (BEST 3) Read eac problem carefully.

More information

THE IDEA OF DIFFERENTIABILITY FOR FUNCTIONS OF SEVERAL VARIABLES Math 225

THE IDEA OF DIFFERENTIABILITY FOR FUNCTIONS OF SEVERAL VARIABLES Math 225 THE IDEA OF DIFFERENTIABILITY FOR FUNCTIONS OF SEVERAL VARIABLES Mat 225 As we ave seen, te definition of derivative for a Mat 111 function g : R R and for acurveγ : R E n are te same, except for interpretation:

More information

CS522 - Partial Di erential Equations

CS522 - Partial Di erential Equations CS5 - Partial Di erential Equations Tibor Jánosi April 5, 5 Numerical Di erentiation In principle, di erentiation is a simple operation. Indeed, given a function speci ed as a closed-form formula, its

More information

Domination Problems in Nowhere-Dense Classes of Graphs

Domination Problems in Nowhere-Dense Classes of Graphs LIPIcs Leibniz International Proceedings in Informatics Domination Problems in Nowere-Dense Classes of Graps Anuj Dawar 1, Stepan Kreutzer 2 1 University of Cambridge Computer Lab, U.K. anuj.dawar@cl.cam.ac.uk

More information

Volume 29, Issue 3. Existence of competitive equilibrium in economies with multi-member households

Volume 29, Issue 3. Existence of competitive equilibrium in economies with multi-member households Volume 29, Issue 3 Existence of competitive equilibrium in economies wit multi-member ouseolds Noriisa Sato Graduate Scool of Economics, Waseda University Abstract Tis paper focuses on te existence of

More information

Introduction to Derivatives

Introduction to Derivatives Introduction to Derivatives 5-Minute Review: Instantaneous Rates and Tangent Slope Recall te analogy tat we developed earlier First we saw tat te secant slope of te line troug te two points (a, f (a))

More information

MATH1151 Calculus Test S1 v2a

MATH1151 Calculus Test S1 v2a MATH5 Calculus Test 8 S va January 8, 5 Tese solutions were written and typed up by Brendan Trin Please be etical wit tis resource It is for te use of MatSOC members, so do not repost it on oter forums

More information

Robotic manipulation project

Robotic manipulation project Robotic manipulation project Bin Nguyen December 5, 2006 Abstract Tis is te draft report for Robotic Manipulation s class project. Te cosen project aims to understand and implement Kevin Egan s non-convex

More information

1 Calculus. 1.1 Gradients and the Derivative. Q f(x+h) f(x)

1 Calculus. 1.1 Gradients and the Derivative. Q f(x+h) f(x) Calculus. Gradients and te Derivative Q f(x+) δy P T δx R f(x) 0 x x+ Let P (x, f(x)) and Q(x+, f(x+)) denote two points on te curve of te function y = f(x) and let R denote te point of intersection of

More information

232 Calculus and Structures

232 Calculus and Structures 3 Calculus and Structures CHAPTER 17 JUSTIFICATION OF THE AREA AND SLOPE METHODS FOR EVALUATING BEAMS Calculus and Structures 33 Copyrigt Capter 17 JUSTIFICATION OF THE AREA AND SLOPE METHODS 17.1 THE

More information

ERROR BOUNDS FOR THE METHODS OF GLIMM, GODUNOV AND LEVEQUE BRADLEY J. LUCIER*

ERROR BOUNDS FOR THE METHODS OF GLIMM, GODUNOV AND LEVEQUE BRADLEY J. LUCIER* EO BOUNDS FO THE METHODS OF GLIMM, GODUNOV AND LEVEQUE BADLEY J. LUCIE* Abstract. Te expected error in L ) attimet for Glimm s sceme wen applied to a scalar conservation law is bounded by + 2 ) ) /2 T

More information

Numerical Differentiation

Numerical Differentiation Numerical Differentiation Finite Difference Formulas for te first derivative (Using Taylor Expansion tecnique) (section 8.3.) Suppose tat f() = g() is a function of te variable, and tat as 0 te function

More information

Section 2: The Derivative Definition of the Derivative

Section 2: The Derivative Definition of the Derivative Capter 2 Te Derivative Applied Calculus 80 Section 2: Te Derivative Definition of te Derivative Suppose we drop a tomato from te top of a 00 foot building and time its fall. Time (sec) Heigt (ft) 0.0 00

More information

Solution. Solution. f (x) = (cos x)2 cos(2x) 2 sin(2x) 2 cos x ( sin x) (cos x) 4. f (π/4) = ( 2/2) ( 2/2) ( 2/2) ( 2/2) 4.

Solution. Solution. f (x) = (cos x)2 cos(2x) 2 sin(2x) 2 cos x ( sin x) (cos x) 4. f (π/4) = ( 2/2) ( 2/2) ( 2/2) ( 2/2) 4. December 09, 20 Calculus PracticeTest s Name: (4 points) Find te absolute extrema of f(x) = x 3 0 on te interval [0, 4] Te derivative of f(x) is f (x) = 3x 2, wic is zero only at x = 0 Tus we only need

More information

5 Ordinary Differential Equations: Finite Difference Methods for Boundary Problems

5 Ordinary Differential Equations: Finite Difference Methods for Boundary Problems 5 Ordinary Differential Equations: Finite Difference Metods for Boundary Problems Read sections 10.1, 10.2, 10.4 Review questions 10.1 10.4, 10.8 10.9, 10.13 5.1 Introduction In te previous capters we

More information

Spike train entropy-rate estimation using hierarchical Dirichlet process priors

Spike train entropy-rate estimation using hierarchical Dirichlet process priors publised in: Advances in Neural Information Processing Systems 26 (23), 276 284. Spike train entropy-rate estimation using ierarcical Diriclet process priors Karin Knudson Department of Matematics kknudson@mat.utexas.edu

More information

LIMITS AND DERIVATIVES CONDITIONS FOR THE EXISTENCE OF A LIMIT

LIMITS AND DERIVATIVES CONDITIONS FOR THE EXISTENCE OF A LIMIT LIMITS AND DERIVATIVES Te limit of a function is defined as te value of y tat te curve approaces, as x approaces a particular value. Te limit of f (x) as x approaces a is written as f (x) approaces, as

More information

2.3 Algebraic approach to limits

2.3 Algebraic approach to limits CHAPTER 2. LIMITS 32 2.3 Algebraic approac to its Now we start to learn ow to find its algebraically. Tis starts wit te simplest possible its, and ten builds tese up to more complicated examples. Fact.

More information

Continuity and Differentiability

Continuity and Differentiability Continuity and Dierentiability Tis capter requires a good understanding o its. Te concepts o continuity and dierentiability are more or less obvious etensions o te concept o its. Section - INTRODUCTION

More information

Material for Difference Quotient

Material for Difference Quotient Material for Difference Quotient Prepared by Stepanie Quintal, graduate student and Marvin Stick, professor Dept. of Matematical Sciences, UMass Lowell Summer 05 Preface Te following difference quotient

More information

The cluster problem in constrained global optimization

The cluster problem in constrained global optimization Te cluster problem in constrained global optimization Te MIT Faculty as made tis article openly available. Please sare ow tis access benefits you. Your story matters. Citation As Publised Publiser Kannan,

More information

THE STURM-LIOUVILLE-TRANSFORMATION FOR THE SOLUTION OF VECTOR PARTIAL DIFFERENTIAL EQUATIONS. L. Trautmann, R. Rabenstein

THE STURM-LIOUVILLE-TRANSFORMATION FOR THE SOLUTION OF VECTOR PARTIAL DIFFERENTIAL EQUATIONS. L. Trautmann, R. Rabenstein Worksop on Transforms and Filter Banks (WTFB),Brandenburg, Germany, Marc 999 THE STURM-LIOUVILLE-TRANSFORMATION FOR THE SOLUTION OF VECTOR PARTIAL DIFFERENTIAL EQUATIONS L. Trautmann, R. Rabenstein Lerstul

More information