Dynamic Programming for Linear-Time Incremental Parsing

Total Page:16

File Type:pdf, Size:1020Kb

Dynamic Programming for Linear-Time Incremental Parsing Dynamic Programming for Linear-Time Incremental Parsing Liang Huang Kenji Sagae USC Information Sciences Institute USC Institute for Creative Technologies 4676 Admiralty Way, Suite 1001 13274 Fiji Way Marina del Rey, CA 90292 Marina del Rey, CA 90292 [email protected] [email protected] Abstract that runs in (almost) linear-time, yet searches over a huge space with dynamic programming? Incremental parsing techniques such as Theoretically, the answer is negative, as Lee shift-reduce have gained popularity thanks (2002) shows that context-free parsing can be used to their efficiency, but there remains a to compute matrix multiplication, where sub-cubic major problem: the search is greedy and algorithms are largely impractical. only explores a tiny fraction of the whole We instead propose a dynamic programming al- space (even with beam search) as op- ogorithm for shift-reduce parsing which runs in posed to dynamic programming. We show polynomial time in theory, but linear-time (with that, surprisingly, dynamic programming beam search) in practice. The key idea is to merge is in fact possible for many shift-reduce equivalent stacks according to feature functions, parsers, by merging “equivalent” stacks inspired by Earley parsing (Earley, 1970; Stolcke, based on feature values. Empirically, our 1995) and generalized LR parsing (Tomita, 1991). algorithm yields up to a five-fold speedup However, our formalism is more flexible and our over a state-of-the-art shift-reduce depen- algorithm more practical. Specifically, we make dency parser with no loss in accuracy. Bet- the following contributions: ter search also leads to better learning, and our final parser outperforms all previously • theoretically, we show that for a large class reported dependency parsers for English of modern shift-reduce parsers, dynamic pro- and Chinese, yet is much faster. gramming is in fact possible and runs in poly- nomial time as long as the feature functions 1 Introduction are bounded and monotonic (which almost al- ways holds in practice); In terms of search strategy, most parsing al- gorithms in current use for data-driven parsing • practically, dynamic programming is up to can be divided into two broad categories: dy- five times faster (with the same accuracy) as namic programming which includes the domi- conventional beam-search on top of a state- nant CKY algorithm, and greedy search which in- of-the-art shift-reduce dependency parser; cludes most incremental parsing methods such as • as a by-product, dynamic programming can shift-reduce.1 Both have pros and cons: the for- output a forest encoding exponentially many mer performs an exact search (in cubic time) over trees, out of which we can draw better and an exponentially large space, while the latter is longer k-best lists than beam search can; much faster (in linear-time) and is psycholinguis- tically motivated (Frazier and Rayner, 1982), but • finally, better and faster search also leads to its greedy nature may suffer from severe search er- better and faster learning. Our final parser rors, as it only explores a tiny fraction of the whole achieves the best (unlabeled) accuracies that space even with a beam. we are aware of in both English and Chi- Can we combine the advantages of both ap- nese among dependency parsers trained on proaches, that is, construct an incremental parser the Penn Treebanks. Being linear-time, it is 1McDonald et al. (2005b) is a notable exception: the MST also much faster than most other parsers, algorithm is exact search but not dynamic programming. even with a pure Python implementation. input: w0 ...wn−1 input: “I saw Al with Joe” step action stack queue axiom 0: h0, ǫi: 0 0 - I ... ℓ : hj, Si : c 1 sh I saw ... sh 2 sh I saw Al ... j<n x ℓ +1: hj + 1, S|wji : c + ξ 3 rex I saw Al ... 4 sh Ixsaw Al with ... x y ℓ : hj, S|s1|s0i : c 5a rey I saw Al with ... re x x 5b sh Ixsaw Al with Joe ℓ +1: hj, S|s1 s0i : c + λ Figure 2: A trace of vanilla shift-reduce. After ℓ : hj, S|s1|s0i : c rey y step (4), the parser branches off into (5a) or (5b). ℓ +1: hj, S|s1 s0i : c + ρ goal 2n − 1: hn, s0i: c queue head position (current word q0 is wj). At each step, we choose one of the three actions: where ℓ is the step, c is the cost, and the shift cost ξ and reduce costs λ and ρ are: 1. sh: move the head of queue, wj, onto stack S as a singleton tree; ξ = w · fsh(j, S) (1) re w f 2. x: combine the top two trees on the stack, λ = · rex (j, S|s1|s0) (2) x s0 and s1, and replace them with tree s1 s0. ρ = w · frey (j, S|s1|s0) (3) 3. rey: combine the top two trees on the stack, y s0 and s1, and replace them with tree s1 s0. Figure 1: Deductive system of vanilla shift-reduce. Note that the shorthand notation txt′ denotes a new tree by “attaching tree t as the leftmost child For convenience of presentation and experimen- of the root of tree t′”. This procedure can be sum- tation, we will focus on shift-reduce parsing for marized as a deductive system in Figure 1. States dependency structures in the remainder of this pa- are organized according to step ℓ, which denotes per, though our formalism and algorithm can also the number of actions accumulated. The parser be applied to phrase-structure parsing. runs in linear-time as there are exactly 2n−1 steps for a sentence of n words. 2 Shift-Reduce Parsing As an example, consider the sentence “I saw Al 2.1 Vanilla Shift-Reduce with Joe” in Figure 2. At step (4), we face a shift- reduce conflict: either combine “saw” and “Al” in Shift-reduce parsing performs a left-to-right scan a rey action (5a), or shift “with” (5b). To resolve of the input sentence, and at each step, choose one this conflict, there is a cost c associated with each of the two actions: either shift the current word state so that we can pick the best one (or few, with onto the stack, or reduce the top two (or more) a beam) at each step. Costs are accumulated in items at the end of the stack (Aho and Ullman, each step: as shown in Figure 1, actions sh, rex, 1972). To adapt it to dependency parsing, we split and rey have their respective costs ξ, λ, and ρ, re re the reduce action into two cases, x and y, de- which are dot-products of the weights w and fea- pending on which one of the two items becomes tures extracted from the state and the action. the head after reduction. This procedure is known as “arc-standard” (Nivre, 2004), and has been en- 2.2 Features gineered to achieve state-of-the-art parsing accu- We view features as “abstractions” or (partial) ob- racy in Huang et al. (2009), which is also the ref- servations of the current state, which is an im- 2 erence parser in our experiments. portant intuition for the development of dynamic More formally, we describe a parser configura- programming in Section 3. Feature templates tion by a state hj, Si where S is a stack of trees are functions that draw information from the fea- s0, s1,... where s0 is the top tree, and j is the ture window (see Tab. 1(b)), consisting of the top few trees on the stack and the first few 2There is another popular variant, “arc-eager” (Nivre, 2004; Zhang and Clark, 2008), which is more complicated words on the queue. For example, one such fea- and less similar to the classical shift-reduce algorithm. ture templatef100 = s0.w ◦ q0.t is a conjunction (a) Features Templates f j,S q w + of two atomic features s0.w and q0.t, capturing ( ) i = j i (1) s0.w s0.t s0.w ◦ s0.t the root word of the top tree s0 on the stack, and s1.w s1.t s1.w ◦ s1.t the part-of-speech tag of the current head word q0 q0.w q0.t q0.w ◦ q0.t w w t t on the queue. See Tab. 1(a) for the list of feature (2) s0. ◦ s1. s0. ◦ s1. s0.t ◦ q0.t s0.w ◦ s0.t ◦ s1.t templates used in the full model. Feature templates s0.t ◦ s1.w ◦ s1.t s0.w ◦ s1.w ◦ s1.t are instantiated for a specific state. For example, at s0.w ◦ s0.t ◦ s1.w s0.w ◦ s0.t ◦ s1 ◦ s1.t (3) s0.t ◦ q0.t ◦ q1.t s1.t ◦ s0.t ◦ q0.t step (4) in Fig. 2, the above template f100 will gen- s0.w ◦ q0.t ◦ q1.t s1.t ◦ s0.w ◦ q0.t erate a feature instance (4) s1.t ◦ s1.lc.t ◦ s0.t s1.t ◦ s1.rc.t ◦ s0.t s1.t ◦ s0.t ◦ s0.rc.t s1.t ◦ s1.lc.t ◦ s0 s0.w Al q0.t IN . ( = ) ◦ ( = ) s1.t ◦ s1.rc.t ◦ s0.w s1.t ◦ s0.w ◦ s0.lc.t (5) s2.t ◦ s1.t ◦ s0.t More formally, we denote f to be the feature func- tion, such that f(j, S) returns a vector of feature (b) ← stack queue → ... q0 q1 ... instances for state hj, Si. To decide which action s2 s1 s0 is the best for the current state, we perform a three- ..
Recommended publications
  • Metaheuristics1
    METAHEURISTICS1 Kenneth Sörensen University of Antwerp, Belgium Fred Glover University of Colorado and OptTek Systems, Inc., USA 1 Definition A metaheuristic is a high-level problem-independent algorithmic framework that provides a set of guidelines or strategies to develop heuristic optimization algorithms (Sörensen and Glover, To appear). Notable examples of metaheuristics include genetic/evolutionary algorithms, tabu search, simulated annealing, and ant colony optimization, although many more exist. A problem-specific implementation of a heuristic optimization algorithm according to the guidelines expressed in a metaheuristic framework is also referred to as a metaheuristic. The term was coined by Glover (1986) and combines the Greek prefix meta- (metá, beyond in the sense of high-level) with heuristic (from the Greek heuriskein or euriskein, to search). Metaheuristic algorithms, i.e., optimization methods designed according to the strategies laid out in a metaheuristic framework, are — as the name suggests — always heuristic in nature. This fact distinguishes them from exact methods, that do come with a proof that the optimal solution will be found in a finite (although often prohibitively large) amount of time. Metaheuristics are therefore developed specifically to find a solution that is “good enough” in a computing time that is “small enough”. As a result, they are not subject to combinatorial explosion – the phenomenon where the computing time required to find the optimal solution of NP- hard problems increases as an exponential function of the problem size. Metaheuristics have been demonstrated by the scientific community to be a viable, and often superior, alternative to more traditional (exact) methods of mixed- integer optimization such as branch and bound and dynamic programming.
    [Show full text]
  • Lecture 4 Dynamic Programming
    1/17 Lecture 4 Dynamic Programming Last update: Jan 19, 2021 References: Algorithms, Jeff Erickson, Chapter 3. Algorithms, Gopal Pandurangan, Chapter 6. Dynamic Programming 2/17 Backtracking is incredible powerful in solving all kinds of hard prob- lems, but it can often be very slow; usually exponential. Example: Fibonacci numbers is defined as recurrence: 0 if n = 0 Fn =8 1 if n = 1 > Fn 1 + Fn 2 otherwise < ¡ ¡ > A direct translation in:to recursive program to compute Fibonacci number is RecFib(n): if n=0 return 0 if n=1 return 1 return RecFib(n-1) + RecFib(n-2) Fibonacci Number 3/17 The recursive program has horrible time complexity. How bad? Let's try to compute. Denote T(n) as the time complexity of computing RecFib(n). Based on the recursion, we have the recurrence: T(n) = T(n 1) + T(n 2) + 1; T(0) = T(1) = 1 ¡ ¡ Solving this recurrence, we get p5 + 1 T(n) = O(n); = 1.618 2 So the RecFib(n) program runs at exponential time complexity. RecFib Recursion Tree 4/17 Intuitively, why RecFib() runs exponentially slow. Problem: redun- dant computation! How about memorize the intermediate computa- tion result to avoid recomputation? Fib: Memoization 5/17 To optimize the performance of RecFib, we can memorize the inter- mediate Fn into some kind of cache, and look it up when we need it again. MemFib(n): if n = 0 n = 1 retujrjn n if F[n] is undefined F[n] MemFib(n-1)+MemFib(n-2) retur n F[n] How much does it improve upon RecFib()? Assuming accessing F[n] takes constant time, then at most n additions will be performed (we never recompute).
    [Show full text]
  • Dynamic Programming Via Static Incrementalization 1 Introduction
    Dynamic Programming via Static Incrementalization Yanhong A. Liu and Scott D. Stoller Abstract Dynamic programming is an imp ortant algorithm design technique. It is used for solving problems whose solutions involve recursively solving subproblems that share subsubproblems. While a straightforward recursive program solves common subsubproblems rep eatedly and of- ten takes exp onential time, a dynamic programming algorithm solves every subsubproblem just once, saves the result, reuses it when the subsubproblem is encountered again, and takes p oly- nomial time. This pap er describ es a systematic metho d for transforming programs written as straightforward recursions into programs that use dynamic programming. The metho d extends the original program to cache all p ossibly computed values, incrementalizes the extended pro- gram with resp ect to an input increment to use and maintain all cached results, prunes out cached results that are not used in the incremental computation, and uses the resulting in- cremental program to form an optimized new program. Incrementalization statically exploits semantics of b oth control structures and data structures and maintains as invariants equalities characterizing cached results. The principle underlying incrementalization is general for achiev- ing drastic program sp eedups. Compared with previous metho ds that p erform memoization or tabulation, the metho d based on incrementalization is more powerful and systematic. It has b een implemented and applied to numerous problems and succeeded on all of them. 1 Intro duction Dynamic programming is an imp ortant technique for designing ecient algorithms [2, 44 , 13 ]. It is used for problems whose solutions involve recursively solving subproblems that overlap.
    [Show full text]
  • Automatic Code Generation Using Dynamic Programming Techniques
    ! Automatic Code Generation using Dynamic Programming Techniques MASTERARBEIT zur Erlangung des akademischen Grades Diplom-Ingenieur im Masterstudium INFORMATIK Eingereicht von: Igor Böhm, 0155477 Angefertigt am: Institut für System Software Betreuung: o.Univ.-Prof.Dipl.-Ing. Dr. Dr.h.c. Hanspeter Mössenböck Linz, Oktober 2007 Abstract Building compiler back ends from declarative specifications that map tree structured intermediate representations onto target machine code is the topic of this thesis. Although many tools and approaches have been devised to tackle the problem of automated code generation, there is still room for improvement. In this context we present Hburg, an implementation of a code generator generator that emits compiler back ends from concise tree pattern specifications written in our code generator description language. The language features attribute grammar style specifications and allows for great flexibility with respect to the placement of semantic actions. Our main contribution is to show that these language features can be integrated into automatically generated code generators that perform optimal instruction selection based on tree pattern matching combined with dynamic program- ming. In order to substantiate claims about the usefulness of our language we provide two complete examples that demonstrate how to specify code generators for Risc and Cisc architectures. Kurzfassung Diese Diplomarbeit beschreibt Hburg, ein Werkzeug das aus einer Spezi- fikation des abstrakten Syntaxbaums eines Programms und der Spezifika- tion der gewuns¨ chten Zielmaschine automatisch einen Codegenerator fur¨ diese Maschine erzeugt. Abbildungen zwischen abstrakten Syntaxb¨aumen und einer Zielmaschine werden durch Baummuster definiert. Fur¨ diesen Zweck haben wir eine deklarative Beschreibungssprache entwickelt, die es erm¨oglicht den Baummustern Attribute beizugeben, wodurch diese gleich- sam parametrisiert werden k¨onnen.
    [Show full text]
  • 1 Introduction 2 Dijkstra's Algorithm
    15-451/651: Design & Analysis of Algorithms September 3, 2019 Lecture #3: Dynamic Programming II last changed: August 30, 2019 In this lecture we continue our discussion of dynamic programming, focusing on using it for a variety of path-finding problems in graphs. Topics in this lecture include: • The Bellman-Ford algorithm for single-source (or single-sink) shortest paths. • Matrix-product algorithms for all-pairs shortest paths. • Algorithms for all-pairs shortest paths, including Floyd-Warshall and Johnson. • Dynamic programming for the Travelling Salesperson Problem (TSP). 1 Introduction As a reminder of basic terminology: a graph is a set of nodes or vertices, with edges between some of the nodes. We will use V to denote the set of vertices and E to denote the set of edges. If there is an edge between two vertices, we call them neighbors. The degree of a vertex is the number of neighbors it has. Unless otherwise specified, we will not allow self-loops or multi-edges (multiple edges between the same pair of nodes). As is standard with discussing graphs, we will use n = jV j, and m = jEj, and we will let V = f1; : : : ; ng. The above describes an undirected graph. In a directed graph, each edge now has a direction (and as we said earlier, we will sometimes call the edges in a directed graph arcs). For each node, we can now talk about out-neighbors (and out-degree) and in-neighbors (and in-degree). In a directed graph you may have both an edge from u to v and an edge from v to u.
    [Show full text]
  • Language and Compiler Support for Dynamic Code Generation by Massimiliano A
    Language and Compiler Support for Dynamic Code Generation by Massimiliano A. Poletto S.B., Massachusetts Institute of Technology (1995) M.Eng., Massachusetts Institute of Technology (1995) Submitted to the Department of Electrical Engineering and Computer Science in partial fulfillment of the requirements for the degree of Doctor of Philosophy at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY September 1999 © Massachusetts Institute of Technology 1999. All rights reserved. A u th or ............................................................................ Department of Electrical Engineering and Computer Science June 23, 1999 Certified by...............,. ...*V .,., . .* N . .. .*. *.* . -. *.... M. Frans Kaashoek Associate Pro essor of Electrical Engineering and Computer Science Thesis Supervisor A ccepted by ................ ..... ............ ............................. Arthur C. Smith Chairman, Departmental CommitteA on Graduate Students me 2 Language and Compiler Support for Dynamic Code Generation by Massimiliano A. Poletto Submitted to the Department of Electrical Engineering and Computer Science on June 23, 1999, in partial fulfillment of the requirements for the degree of Doctor of Philosophy Abstract Dynamic code generation, also called run-time code generation or dynamic compilation, is the cre- ation of executable code for an application while that application is running. Dynamic compilation can significantly improve the performance of software by giving the compiler access to run-time infor- mation that is not available to a traditional static compiler. A well-designed programming interface to dynamic compilation can also simplify the creation of important classes of computer programs. Until recently, however, no system combined efficient dynamic generation of high-performance code with a powerful and portable language interface. This thesis describes a system that meets these requirements, and discusses several applications of dynamic compilation.
    [Show full text]
  • Bellman-Ford Algorithm
    The many cases offinding shortest paths Dynamic programming Bellman-Ford algorithm We’ve already seen how to calculate the shortest path in an Tyler Moore unweighted graph (BFS traversal) We’ll now study how to compute the shortest path in different CSE 3353, SMU, Dallas, TX circumstances for weighted graphs Lecture 18 1 Single-source shortest path on a weighted DAG 2 Single-source shortest path on a weighted graph with nonnegative weights (Dijkstra’s algorithm) 3 Single-source shortest path on a weighted graph including negative weights (Bellman-Ford algorithm) Some slides created by or adapted from Dr. Kevin Wayne. For more information see http://www.cs.princeton.edu/~wayne/kleinberg-tardos. Some code reused from Python Algorithms by Magnus Lie Hetland. 2 / 13 �������������� Shortest path problem. Given a digraph ����������, with arbitrary edge 6. DYNAMIC PROGRAMMING II weights or costs ���, find cheapest path from node � to node �. ‣ sequence alignment ‣ Hirschberg's algorithm ‣ Bellman-Ford 1 -3 3 5 ‣ distance vector protocols 4 12 �������� 0 -5 ‣ negative cycles in a digraph 8 7 7 2 9 9 -1 1 11 ����������� 5 5 -3 13 4 10 6 ������������� ���������������������������������� 22 3 / 13 4 / 13 �������������������������������� ��������������� Dijkstra. Can fail if negative edge weights. Def. A negative cycle is a directed cycle such that the sum of its edge weights is negative. s 2 u 1 3 v -8 w 5 -3 -3 Reweighting. Adding a constant to every edge weight can fail. 4 -4 u 5 5 2 2 ���������������������� c(W) = ce < 0 t s e W 6 6 �∈ 3 3 0 v -3 w 23 24 5 / 13 6 / 13 ���������������������������������� ���������������������������������� Lemma 1.
    [Show full text]
  • Notes on Dynamic Programming Algorithms & Data Structures 1
    Notes on Dynamic Programming Algorithms & Data Structures Dr Mary Cryan These notes are to accompany lectures 10 and 11 of ADS. 1 Introduction The technique of Dynamic Programming (DP) could be described “recursion turned upside-down”. However, it is not usually used as an alternative to recursion. Rather, dynamic programming is used (if possible) for cases when a recurrence for an algorithmic problem will not run in polynomial-time if it is implemented recursively. So in fact Dynamic Programming is a more- powerful technique than basic Divide-and-Conquer. Designing, Analysing and Implementing a dynamic programming algorithm is (like Divide- and-Conquer) highly problem specific. However, there are particular features shared by most dynamic programming algorithms, and we describe them below on page 2 (dp1(a), dp1(b), dp2, dp3). It will be helpful to carry along an introductory example-problem to illustrate these fea- tures. The introductory problem will be the problem of computing the nth Fibonacci number, where F(n) is defined as follows: F0 = 0; F1 = 1; Fn = Fn-1 + Fn-2 (for n ≥ 2). Since Fibonacci numbers are defined recursively, the definition suggests a very natural recursive algorithm to compute F(n): Algorithm REC-FIB(n) 1. if n = 0 then 2. return 0 3. else if n = 1 then 4. return 1 5. else 6. return REC-FIB(n - 1) + REC-FIB(n - 2) First note that were we to implement REC-FIB, we would not be able to use the Master Theo- rem to analyse its running-time. The recurrence for the running-time TREC-FIB(n) that we would get would be TREC-FIB(n) = TREC-FIB(n - 1) + TREC-FIB(n - 2) + Θ(1); (1) where the Θ(1) comes from the fact that, at most, we have to do enough work to add two values together (on line 6).
    [Show full text]
  • Parallel Technique for the Metaheuristic Algorithms Using Devoted Local Search and Manipulating the Solutions Space
    applied sciences Article Parallel Technique for the Metaheuristic Algorithms Using Devoted Local Search and Manipulating the Solutions Space Dawid Połap 1,* ID , Karolina K˛esik 1, Marcin Wo´zniak 1 ID and Robertas Damaševiˇcius 2 ID 1 Institute of Mathematics, Silesian University of Technology, Kaszubska 23, 44-100 Gliwice, Poland; [email protected] (K.K.); [email protected] (M.W.) 2 Department of Software Engineering, Kaunas University of Technology, Studentu 50, LT-51368, Kaunas, Lithuania; [email protected] * Correspondence: [email protected] Received: 16 December 2017; Accepted: 13 February 2018 ; Published: 16 February 2018 Abstract: The increasing exploration of alternative methods for solving optimization problems causes that parallelization and modification of the existing algorithms are necessary. Obtaining the right solution using the meta-heuristic algorithm may require long operating time or a large number of iterations or individuals in a population. The higher the number, the longer the operation time. In order to minimize not only the time, but also the value of the parameters we suggest three proposition to increase the efficiency of classical methods. The first one is to use the method of searching through the neighborhood in order to minimize the solution space exploration. Moreover, task distribution between threads and CPU cores can affect the speed of the algorithm and therefore make it work more efficiently. The second proposition involves manipulating the solutions space to minimize the number of calculations. In addition, the third proposition is the combination of the previous two. All propositions has been described, tested and analyzed due to the use of various test functions.
    [Show full text]
  • Branch and Bound, Divide and Conquer, and More TA: Chinmay Nirkhe ([email protected]) & Catherine Ma ([email protected])
    CS 38 Introduction to Algorithms Week 7 Recitation Notes Branch and Bound, Divide and Conquer, and More TA: Chinmay Nirkhe ([email protected]) & Catherine Ma ([email protected]) 1 Branch and Bound 1.1 Preliminaries So far we have covered Dynamic Programming, Greedy Algorithms, and (some) Graph Algorithms. Other than a few examples with Knapsack and Travelling Salesman, however, we have mostly covered P algorithms. Now we turn to look at some NP algorithms. Recognize, that, we are not going to come up with any P algorithms for these problems but we will be able to radically improve runtime over a na¨ıve algorithm. Specifically we are going to discuss a technique called Branch and Bound. Let's first understand the intu- ition behind the technique. Assume we are trying to maximize a function f over a exponentially large set X. However, not only is the set exponentially large but f might be computationally intensive on this set. Fortunately, we know of a function h such that f ≤ h everywhere. Our na¨ıve strategy is to keep a maximum value m (initially at −∞) and for each x 2 X, update it by m maxfm; f(x)g. However, an alternative strategy would be to calculate h(x) first. If h(x) ≤ m then we know f(x) ≤ h(x) ≤ m so the maximum will not change. So we can effectively avoid computing f(x) saving us a lot on computation! However, if h(x) > m then we cannot tell anything about f(x) so we would have to compute f(x) to check the maximum.
    [Show full text]
  • Toward a Model for Backtracking and Dynamic Programming
    TOWARD A MODEL FOR BACKTRACKING AND DYNAMIC PROGRAMMING Michael Alekhnovich, Allan Borodin, Joshua Buresh-Oppenheim, Russell Impagliazzo, Avner Magen, and Toniann Pitassi Abstract. We propose a model called priority branching trees (pBT ) for backtrack- ing and dynamic programming algorithms. Our model generalizes both the priority model of Borodin, Nielson and Rackoff, as well as a simple dynamic programming model due to Woeginger, and hence spans a wide spectrum of algorithms. After witnessing the strength of the model, we then show its limitations by providing lower bounds for algorithms in this model for several classical problems such as Interval Scheduling, Knapsack and Satisfiability. Keywords. Greedy Algorithms, Dynamic Programming, Models of Computation, Lower Bounds. Subject classification. 68Q10 1. Introduction The “Design and Analysis of Algorithms” is a basic component of the Computer Science Curriculum. Courses and texts for this topic are often organized around a toolkit of al- gorithmic paradigms or meta-algorithms such as greedy algorithms, divide and conquer, dynamic programming, local search, etc. Surprisingly (as this is often the main “theory course”), these algorithmic paradigms are rarely, if ever, precisely defined. Instead, we provide informal definitional statements followed by (hopefully) well chosen illustrative examples. Our informality in algorithm design should be compared to computability the- ory where we have a well accepted formalization for the concept of an algorithm, namely that provided by Turing machines and its many equivalent computational models (i.e. consider the almost universal acceptance of the Church-Turing thesis). While quantum computation may challenge the concept of “efficient algorithm”, the benefit of having a well defined concept of an algorithm and a computational step is well appreciated.
    [Show full text]
  • Greedy Algorithms CLRS 16.1-16.2
    Greedy Algorithms CLRS 16.1-16.2 Today we discuss a technique called “greedy”. To understand the greedy technique, it’s best to understand the differences between greedy and dynamic programming. Dynamic programming: • The problem must have the optimal substructure property: the optimal solution to the problem contains within it optimal solutions to subproblems. This allows for a recursive solution. • The idea of dynamic programming is the following: At every step , evaluate all choices recursively and pick the best. A choice is evaluated recursively, meaning all its choices are evaluated and all their choices are evaluated, and so on; basically to evaluate a choice, you have to go through the whole recursion tree for that choice. • The table: The subproblems are “tabulated”, that is, they are stored in a table; using the table to store solutions to subproblems means that subproblem are computed only once. Typically the number of different subproblems is polynomial, but the recursive algorithm implementing the recursive formulation of the problem is exponential. This is because of overlapping calls to same subproblem. So if you don’t use the table to cache partial solutions, you incur a significant penalty (Often the difference is polynomial vs. exponential). • Sometimes we might want to take dynamic programming a step further, and eliminate the recursion —- this is purely for eliminating the overhead of recursion, and does not change the Θ() of the running time. We can think of dynamic programming as filling the sub-problems solutions table bottom-up, without recursion. The greedy technique: • As with dynamic programming, in order to be solved with the greedy technique, the problem must have the optimal substructure property • The problems that can be solved with the greedy method are a subset of those that can be solved with dynamic programming.
    [Show full text]