Monte-Carlo Tree Search Using Expert Knowledge: an Application to Computer Go and Human Genetics

Total Page:16

File Type:pdf, Size:1020Kb

Monte-Carlo Tree Search Using Expert Knowledge: an Application to Computer Go and Human Genetics Curso 2012/13 CIENCIAS Y TECNOLOGÍAS/23 I.S.B.N.: 978-84-15910-90-9 SANTIAGO BASALDÚA LEMARCHAND Monte-Carlo tree search using expert knowledge: an application to computer go and human genetics Directores J. MARCOS MORENO VEGA CARLOS A. FLORES INFANTE SOPORTES AUDIOVISUALES E INFORMÁTICOS Serie Tesis Doctorales ciencias 23 (Santiago Basaldúa Lemarchand).indd 1 18/02/2014 11:24:43 Universidad de La Laguna Abstract Monte-Carlo Tree Search Using Expert Knowledge: An Application to Computer Go and Human Genetics During the years in which the research described in this PhD dissertation was done, Monte-Carlo Tree Search has become the preeminent algorithm in many AI and computer science fields. This dissertation analyzes how expert knowledge and also online learned knowledge can be used to enhance the search. The work describes two different implementations: as a two player search in computer go and as an optimization method in human genetics. It is established that in large problems MCTS has to be combined with domain specific or online learned knowledge to improve its strength. This work analyzes different successful ideas about how to do it, the resulting findings and their implications, hence improving our insight of MCTS. The main contributions to the field are: an analytical mathematical model improving the understanding of simulations, a problem definition and a framework including code and data to compare algorithms in human genetics and three successful implementations: in the field of 19x19 go openings named M-eval, in the field of learning playouts and in the field of genetic etiology. Also, an open source integer representation of proportions as Win/Loss States (WLS), a negative result in the field of playouts, an unexpected finding of a possible problem in optimization and further insight on the limitations of MCTS are worth mentioning. With the exception of some background materials introducing the game of go and MCTS, the dissertation is entirely written in English. Incorporación de conocimiento experto a la búsqueda en árbol mediante procesos estocásticos y su aplicación en el juego abstracto del go y en genética humana Durante el periodo de desarrollo de esta tesis Monte-Carlo Tree Search (MCTS la búsqueda en árbol mediante procesos estocásticos) se ha convertido en el algoritmo principal en muchos problemas de inteligencia artificial e informática. Esta tesis analiza la incorporación de conocimiento experto para mejorar la búsqueda. El trabajo describe dos aplicaciones: una en el juego del go por ordenador y otra en el campo de la genética humana. Es un hecho establecido que, en problemas complejos, MCTS requiere del apoyo de conocimiento específico o aprendido online para mejorar su rendimiento. Lo que este trabajo analiza son diferentes ideas de cómo hacerlo, sus resultados e implicaciones, mejorando así nuestra comprensión de MCTS. Las principales contribuciones al área son: un modelo analítico de las simulaciones que mejora la comprensión del papel de las simulaciones, un marco competitivo incluyendo código y datos para comparar métodos en etiología genética y tres aplicaciones con éxito: una en el campo de las aperturas en go de 19x19 llamada M-eval, otra sobre simulaciones que aprenden y una en etiología genética. Además, merece la pena destacar: un modelo para representar proporciones mediante estados llamado WLS con software libre, un resultado negativo sobre una idea para las simulaciones, el descubrimiento inesperado de un posible problema utilizando MCTS en optimización y un análisis original de las limitaciones. Con la excepción de unos materiales introductorios sobre el juego del go y Monte-Carlo Tree Search, la tesis está completamente escrita en inglés. by Santiago Basaldúa Lemarchand TABLE OF CONTENTS CHAPTER 1. INTRODUCTION ................................................................................................................. 2 1.1 THESIS OVERVIEW FOR ENGLISH READERS ..................................................................................... 2 1.2 DESCRIPCIÓN DE LA TESIS EN ESPAÑOL ........................................................................................... 3 CHAPTER 2. BACKGROUND (IN SPANISH) .......................................................................................... 5 2.1 INTRODUCCIÓN AL JUEGO DEL GO ................................................................................................... 5 2.1.1 Historia del juego del go ............................................................................................................ 5 2.1.2 Las reglas básicas del go ........................................................................................................... 7 2.1.3 Evaluación de la habilidad del los jugadores ............................................................................12 2.1.4 El papel del conocimiento en go ................................................................................................13 2.2 EL GO COMO PROBLEMA COMPUTACIONAL ....................................................................................14 2.3 INTRODUCCIÓN A MCTS ...............................................................................................................17 2.3.1 UCB y el dilema de exploración/explotación .............................................................................17 2.3.2 Los métodos Monte-Carlo ingenuos ..........................................................................................18 2.3.3 UCT es UCB aplicado en un árbol a una evaluación estocástica ...............................................18 2.3.4 Ejemplo: UCT aplicado al SPP .................................................................................................20 2.3.5 MCTS .......................................................................................................................................25 2.4 APLICACIÓN DE MCTS EN GENÉTICA HUMANA ..............................................................................26 2.4.1 Breve descripción del problema ................................................................................................26 2.4.2 El algoritmo de clasificación utilizado ......................................................................................27 2.4.3 Las medidas de asociación comúnmente empleadas ..................................................................27 2.4.4 Los experimentos realizados......................................................................................................28 2.4.5 Los resultados obtenidos ...........................................................................................................29 CHAPTER 3. APPLYING KNOWLEDGE IN THE TREE SEARCH .....................................................30 3.1 SUCCESSFUL IDEAS USED IN OTHER PROGRAMS .............................................................................30 3.1.1 Biasing moves ...........................................................................................................................31 3.1.2 RAVE ........................................................................................................................................31 3.1.3 Progressive widening ................................................................................................................32 3.1.4 Parallelization of MCTS ............................................................................................................34 3.2 INCLUDING A PRIORI INFORMATION IN THE TREE............................................................................35 3.2.1 Mathematical models described to deal with a priori information ..............................................36 3.2.2 A priori information and tree search as implemented in GoKnot ...............................................38 3.3 CURRENT LIMITS OF MCTS ...........................................................................................................39 3.3.1 Limits in terms of computer science ...........................................................................................39 3.3.2 Limits in go terms......................................................................................................................40 3.3.3 The opening as a limitation .......................................................................................................41 3.3.4 Divide and conquer ...................................................................................................................42 3.4 M-EVAL: A MULTIVARIATE FULL BOARD EVALUATION FUNCTION .................................................44 3.4.1 Positional judgment in go ..........................................................................................................44 3.4.2 M-eval rationale .......................................................................................................................45 3.4.3 NMF: Non-negative matrix factorization ...................................................................................46 3.4.4 Qualities implemented in M-eval ...............................................................................................47 3.4.5 M-eval search ...........................................................................................................................49 i + 3.4.6 Offline learning of H pd ............................................................................................................50 3.4.7 Online learning of L'd1 .............................................................................................................52
Recommended publications
  • Generating Single and Multiple Cooperative Heuristics for the One Dimensional Bin Packing Problem Using a Single Node Genetic Programming Island Model
    Generating Single and Multiple Cooperative Heuristics for the One Dimensional Bin Packing Problem Using a Single Node Genetic Programming Island Model Kevin Sim Emma Hart Institute for Informatics and Digital Innovation Institute for Informatics and Digital Innovation Edinburgh Napier University Edinburgh Napier University Edinburgh, Scotland, UK Edinburgh, Scotland, UK [email protected] [email protected] ABSTRACT exhibit superior performance than similar human designed Novel deterministic heuristics are generated using Single Node heuristics and can be used to produce collections of heuris- Genetic Programming for application to the One Dimen- tics that collectively maximise the potential of selective HH. sional Bin Packing Problem. First a single deterministic The papers contribution is two fold. Single heuristics ca- heuristic was evolved that minimised the total number of pable of outperforming a range of well researched deter- bins used when applied to a set of 685 training instances. ministic heuristics are generated by combining features ex- Following this, a set of heuristics were evolved using a form tracted from those heuristics. Secondly an island model[19] of cooperative co-evolution that collectively minimise the is adapted to use multiple SNGP implementations to gener- number of bins used across the same set of problems. Re- ate diverse sets of heuristics which collectively outperform sults on an unseen test set comprising a further 685 prob- any of the single heuristics when used in isolation. Both ap- lem instances show that the single evolved heuristic out- proaches are trained and evaluated using equal divisions of a performs existing deterministic heuristics described in the large set of 1370 benchmark problem instances sourced from literature.
    [Show full text]
  • Heft 1/2018 93. Jahrgang
    Heft 1/2018 93. Jahrgang 1 DGoZ 1/2018 Inhalt Vorwort Seit langer Zeit hat die DGoZ mal wieder 64+4 Niederländisches Go-Werbeplakat aus den Seiten, also ihre praktisch maximale Dicke. Das 1980er Jahren ......................................Titel liegt einerseits daran, dass die Spielabendliste erst Vorwort, Inhalt ...............................................2 in dieser Ausgabe abgedruckt ist und allein schon Nachrichten ........................................... 2–6 7,5 Seiten lang ist, andererseits an der Fülle von Ausschreibung: DJGM ............................... 7 eingesendeten Berichten, die nur schwer hätten auf Turnierberichte .......................................... 8 spätere Ausgaben verschoben werden können – oder Lee Hajin: Zulassungsgespräch ................... 9 dies schon wurden. Viele lesenswerte Artikel! EGC-2018-Helfersuche ......................10–11 Mit dem Jubiläums-Kidocup wirft ein großes Ein Go-Kongress im Januar? ............... 12–13 Go-Ereignis seinen Schatten voraus, was auch an 28. Internationales Paar -Go-Turnier ........ 14 der DGoZ nicht spurlos vorbei geht: Neben einer Hwang Inseongs Winter-Go-Camp .... 15–16 Ankündigung im Nachrichtenteil hat Yoon Young Interview mit Hwang Inseong ............ 16–17 Das Seidenstraßenturnier .................... 18–19 Sun zwei Partien unserer Kidocup-Gäste Lee Chang- Die EGF Academy .............................. 20–21 ho 9p und Choi Jeong 9p kommentiert. Spannend! Anfängerprobleme ..............................22–23 Besonders hinweisen möchte ich noch auf den Hil-
    [Show full text]
  • Reglas De Congo: Palo Monte Mayombe) a Book by Lydia Cabrera an English Translation from the Spanish
    THE KONGO RULE: THE PALO MONTE MAYOMBE WISDOM SOCIETY (REGLAS DE CONGO: PALO MONTE MAYOMBE) A BOOK BY LYDIA CABRERA AN ENGLISH TRANSLATION FROM THE SPANISH Donato Fhunsu A dissertation submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment of the requirements for the degree of Doctor of Philosophy in the Department of English and Comparative Literature (Comparative Literature). Chapel Hill 2016 Approved by: Inger S. B. Brodey Todd Ramón Ochoa Marsha S. Collins Tanya L. Shields Madeline G. Levine © 2016 Donato Fhunsu ALL RIGHTS RESERVED ii ABSTRACT Donato Fhunsu: The Kongo Rule: The Palo Monte Mayombe Wisdom Society (Reglas de Congo: Palo Monte Mayombe) A Book by Lydia Cabrera An English Translation from the Spanish (Under the direction of Inger S. B. Brodey and Todd Ramón Ochoa) This dissertation is a critical analysis and annotated translation, from Spanish into English, of the book Reglas de Congo: Palo Monte Mayombe, by the Cuban anthropologist, artist, and writer Lydia Cabrera (1899-1991). Cabrera’s text is a hybrid ethnographic book of religion, slave narratives (oral history), and folklore (songs, poetry) that she devoted to a group of Afro-Cubans known as “los Congos de Cuba,” descendants of the Africans who were brought to the Caribbean island of Cuba during the trans-Atlantic Ocean African slave trade from the former Kongo Kingdom, which occupied the present-day southwestern part of Congo-Kinshasa, Congo-Brazzaville, Cabinda, and northern Angola. The Kongo Kingdom had formal contact with Christianity through the Kingdom of Portugal as early as the 1490s.
    [Show full text]
  • 13. Mathematics University of Central Oklahoma
    Southwestern Oklahoma State University SWOSU Digital Commons Oklahoma Research Day Abstracts 2013 Oklahoma Research Day Jan 10th, 12:00 AM 13. Mathematics University of Central Oklahoma Follow this and additional works at: https://dc.swosu.edu/ordabstracts Part of the Animal Sciences Commons, Biology Commons, Chemistry Commons, Computer Sciences Commons, Environmental Sciences Commons, Mathematics Commons, and the Physics Commons University of Central Oklahoma, "13. Mathematics" (2013). Oklahoma Research Day Abstracts. 12. https://dc.swosu.edu/ordabstracts/2013oklahomaresearchday/mathematicsandscience/12 This Event is brought to you for free and open access by the Oklahoma Research Day at SWOSU Digital Commons. It has been accepted for inclusion in Oklahoma Research Day Abstracts by an authorized administrator of SWOSU Digital Commons. An ADA compliant document is available upon request. For more information, please contact [email protected]. Abstracts from the 2013 Oklahoma Research Day Held at the University of Central Oklahoma 05. Mathematics and Science 13. Mathematics 05.13.01 A simplified proof of the Kantorovich theorem for solving equations using scalar telescopic series Ioannis Argyros, Cameron University The Kantorovich theorem is an important tool in Mathematical Analysis for solving nonlinear equations in abstract spaces by approximating a locally unique solution using the popular Newton-Kantorovich method.Many proofs have been given for this theorem using techniques such as the contraction mapping principle,majorizing sequences, recurrent functions and other techniques.These methods are rather long,complicated and not very easy to understand in general by undergraduate students.In the present paper we present a proof using simple telescopic series studied first in a Calculus II class.
    [Show full text]
  • Circuit of Computer Science Unplugged Activities Based on the Life of Ada Lovelace
    Circuit of Computer Science Unplugged activities based on the life of Ada Lovelace 1st Given Name Surname 2nd Given Name Surname 3rd Given Name Surname dept. name of organization (of Aff.) dept. name of organization (of Aff.) dept. name of organization (of Aff.) name of organization (of Aff.) name of organization (of Aff.) name of organization (of Aff.) City, Country City, Country City, Country email address email address email address Abstract—Ada Lovelace’s life is a source of inspiration for defining what are the highest priority tasks and iv) create women and men of all ages, for being a bright-minded and algorithms to solve them. visionary person. His greatest achievement was to have written Unplugged computing occurs through playful activities, the first computer program in history. This article presents a methodological proposal for didactic use of a children’s book group dynamics, and material elaboration. Bell et al. (2015) about Ada Lovelace by proposing a circuit of unplugged activities have been concerned with solving issues that can arise as soon in order to refine the computational thinking (CT) in children and as we are introduced to computing: how do computers work adolescents. This methodology was applied in a workshop during and what do computers solve? The book of such authors, titled the BLIND in the BLIND Conference. A quali-quantitative “CS unplugged: Computer Science without a Computer”, has analysis performed with participants indicates the suitability of the proposed methodology. many activities that can be reproduced with children in an easy Index Terms—computational thinking, cs unplugged, stem, and intuitive way.
    [Show full text]
  • Object-Based Image Analysis for Detecting Landforms Diagnostic of Natural Hazards
    Faculty of Science Institute of Earth and Environmental Science Geohazards Research Group ____________________________________________________________________________ OBJECT-BASED IMAGE ANALYSIS FOR DETECTING LANDFORMS DIAGNOSTIC OF NATURAL HAZARDS Dissertation For the Degree of Doctor of Natural Sciences (Dr. rer. nat.) In the Scientific Discipline of Geohazards University of Potsdam Submitted by Karolina KORZENIOWSKA Potsdam, Germany April 2017 Published online at the Institutional Repository of the University of Potsdam: URN urn:nbn:de:kobv:517-opus4-402240 http://nbn-resolving.de/urn:nbn:de:kobv:517-opus4-402240 Contents Contents ......................................................................................................................................... I Declaration of Authorship .......................................................................................................... III Acknowledgments ....................................................................................................................... V Abstract ...................................................................................................................................... VII Zusammenfassung ...................................................................................................................... IX Abbreviations .............................................................................................................................. XI List of Figures .........................................................................................................................
    [Show full text]
  • Heft 2/2019 94. Jahrgang
    Heft 2/2019 94. Jahrgang 1 DGoZ 2/2019 Inhalt Vorwort Recht spät kommt die zweite Ausgabe des Jahres, Fighting Spirit / Kampfgeist (von Kim Ou- aber sie kommt hoffentlich nicht mehr als vier weleen) ..................................................Titel Wochen zu spät. Die 3/2009 soll dann – wie im- mer – noch vor dem EGC in Brüssel fertiggestellt Vorwort, Inhalt, Nachrichten ......................2–5 werden und die übrigen drei Hefte erscheinen dann Turnierberichte ...................................... 5–9 in der zweiten Jahreshälfte. Es wird allerdings noch jemand für den EGC-Bericht gesucht, da ich selbst Wie viel % kostet ein Zwei-Punkte-Fehler? 10–15 im Sommer alternativ auf Kreta sein werde … Auschreibungen: DDGM, DM-Vorrunde ... 13 Tobias Berben Kinderseite(n) ....................................16–17 Yoon Young Sun kommentiert (45, 1+2) ..18–28 Murugandi-Go-Art Impressum ................................................ 28 Der niederländische Go-Spieler Kim Ouweleen, auch bekannt als Murugandi, ist Autor (2016 Künstliche Intelligenz (2) .................... 29–31 European Go Yearbook und Kamagurkistan) und Pokale ................................................. 32–33 Kunsthistoriker. Er macht auch selbst Kunst und ist freiberuflicher Illustrator. Vor kurzem hat Kim Fernostnachrichten ............................. 34–39 eine Serie von 10 Go-Zeichnungen erstellt, die er Der etwas andere Zug (29) ..................40–46 privat als Postkarten und Poster verkauft. Das Co- ver dieser Ausgabe der Deutschen Go Zeitung ist Go-Probleme .....................................47–48 eines dieser Werke, ebenso wie eine Farbversion Mitgliedsantrag ........................................ 49 des Waschbärhundes, der seinen Bauch trommelt, der auf dem Cover der DGoZ-Ausgabe 1/2015 DGoB-Organe ......................................... 50 zu sehen war. Die Zeichnungen sind Wortspiele Anzeige: Hebsacker Verlag ....................... 51 auf Go-Shapes, die Spiele von AlphaGo und Lee Sedol, der Samurai von Yoshitoshi und vieles Turnierkalender ...........................
    [Show full text]
  • Fujitsu Group Sustainability Report 2008
    2008 FUJITSU GROUP SUST AINABILIT Y REPOR T FUJITSU GROUP SUSTAINABILITY REPORT Legibility Considerations We have reviewed this report using our ColorSelector tool to choose a highly accessible color combination so that the text and f igures will be as legible as possible to the widest range of readers. Consideration for the Environment • It is printed on FSC Certif ied Paper as designated by the Forest Stewardship Council in order to help preserve forestry resources. • It uses vegetable oil inks that do not include volatile organic compounds. • This report has been printed using waterless printing, which reduces the amount of harmful materials used and emitted. FUJITSU LIMITED http://www.fujitsu.com/global/about/environment/ All brand names and product names are trademarks and registered trademarks of their respective holders. AL0052-1-Oct.2008 AP Editorial Policy The 2008 Fujitsu Group Sustainability Report describes the Time Frame thinking, efforts, and accomplishments regarding the social and This report primarily focuses on the thinking, efforts, and accomplishments environmental aspects of Fujitsu Group operations based on the of the Fujitsu Group for fiscal 2007, the period from April 1, 2007 to March 31, 2008, with the data reflecting the actual results for that period. Fujitsu Way, the vision and principles of the Fujitsu Group. The report also includes the Fujitsu Group’s thinking, approaches, and activity data prior to April 1, 2007 and for the period since April 1, 2008. • This report is produced in printed format and published on our website. Organizations Covered •This report presents our basic thinking with respect to the various stakeholders of the Fujitsu Group and the global In principle, this report covers the entire Fujitsu Group, while the environment, concentrating on specific approaches taken and environment-related sections cover a total of 103 consolidated subsidiaries results achieved.
    [Show full text]
  • Lesson Plans for Go in Schools
    Go Lesson Plans 1 Lesson Plans for Go in Schools By Gordon E. Castanza, Ed. D. October 19, 2011 Published By Rittenberg Consulting Group 7806 108th St. NW Gig Harbor, WA 98332 253-853-4831 ©Gordon E. Castanza, Ed. D. 10/19/11 DRAFT Go Lesson Plans 2 Table of Contents Acknowledgements ......................................................................................................................... 4 Purpose/Rationale ........................................................................................................................... 5 Lesson Plan One ............................................................................................................................. 7 Basic Ideas .................................................................................................................................. 7 Introduction ............................................................................................................................... 11 The Puzzle ................................................................................................................................. 13 Surround to Capture .................................................................................................................. 14 First Capture Go ........................................................................................................................ 16 Lesson Plan Two ........................................................................................................................... 19 Units &
    [Show full text]
  • TOM-DOC V3.2A Has the Following New Features: Support of NK2/4 Disks for the Internal Work Files
    Introduction 1 Introduction 1.1 Product Characteristics TOM-DOC edits texts as ready-to-print documents based on commands entered with the text. Editing functions include automatic classification of a text into chapters, sub- chapters, sections, etc., line and page breaks, automatic hyphenation in a number of national languages and automatic generation of table of contents and keyword index. TOM-DOC is particularly well suited for producing documents required for development projects, such as studies, specifications, performance descriptions and manuals. TOM-DOC can be used directly as a separate program or interactively together with the product TOM-TI. When TOM-DOC is called directly, the TOM-DOC texts already entered using the BS2000 editor $EDT are read from system input and the edited result is written to the SYSLST system file. In interactive mode, the texts are entered and modified in the TOM-TI editor and formatted as documents using TOM-DOC. The original text and the formatted document can be displayed at the same time in different windows of the editor. TOM-DOC (SINIX), which is largely compatible with TOM-DOC (BS2000), is available on the SINIX platform. This means that text files can be created and/or edited under either BS2000 or SINIX. TOM-DOC (SINIX) is part of SoftBench-LNK and works with the MAXed/XM (SINIX) editor. U2407-J-Z125-5-7600 1 Introduction Print file Area 0 Input area TOM-DOC Area 1 Editing result TOM-TI work areas Figure 1-1: Interactive operation of TOM-DOC under TOM-TI Description of functions TOM-DOC offers the following functions: • Definition of uniform page layout with header/footer • Definition of variable page formats (DIN A3, DIN A4, ...) • Pagination by chapter • Automatic generation of table of contents and keyword index • Processing of continuous text and fixed text segments • Centering of texts • Automatic classification of texts into chapters, sub-chapters, sections, etc.
    [Show full text]
  • By Prof. Mcgonagall Solution to STEPS TOLD: the Grid with The
    By Prof. McGonagall weeks belong to a set, such that their an- swers form this meta. Those answers are Solution to STEPS TOLD: reproduced below for your convenience. Thegridwiththewhiteandblackcircles Especially now, without a title or flavor- (ignoring the letters for now) is the Nikoli text to give outright hints, it is important logic puzzle ‘‘Masyu’’. Solving that puzzle to keep in mind Rule 9: Always be on the yieldsapaththroughthegrid. Takingeach lookout for themes and patterns. Para- letteralongthatpathspellstheinstruction graph breaks in a block of text, repeated ‘‘MAKE A CRYPTIC CLUE BY USING THE words or letters, apparent typos or non- REMAINING LETTERS IN ORDER’’. standard word selections -- these all can When read left to right, top to bottom, point to clues. Keep your eyes open, and the unused letters in the grid spell the investigate anything that looks unusual. phrase ‘‘SILLY SEATS RING RED PEERS’’. Enjoy your last day of class, and as This is read as a cryptic crossword clue: always, if you think you have the answer, the definition is ‘‘peers’’, and the wordplay submit it on our website below. tells you to mix up the letters of ‘‘seats’’ We’ll see some of you this Sunday at and place them around the letter ‘R’ (for the Berkeley Mystery Hunt! red). This results in the answer, STARES. The title is another wordplay clue to the METAPUZZLE same answer -- ‘‘told’’ indicates a homo- phone of ‘‘stairs’’, a synonym of ‘‘steps’’. DOTING Congratulations to Jevon Heath and REDUB Melinda Fricke, who first submitted the COLORED correct answer to this puzzle. ONES STARTING This now brings us to the metapuz- REDSTAR zle, a new puzzle consisting solely of the COSTARRED answers to all previous puzzles.
    [Show full text]
  • ENDER's GAME by Orson Scott Card Chapter 1 -- Third
    ENDER'S GAME by Orson Scott Card Chapter 1 -- Third "I've watched through his eyes, I've listened through his ears, and tell you he's the one. Or at least as close as we're going to get." "That's what you said about the brother." "The brother tested out impossible. For other reasons. Nothing to do with his ability." "Same with the sister. And there are doubts about him. He's too malleable. Too willing to submerge himself in someone else's will." "Not if the other person is his enemy." "So what do we do? Surround him with enemies all the time?" "If we have to." "I thought you said you liked this kid." "If the buggers get him, they'll make me look like his favorite uncle." "All right. We're saving the world, after all. Take him." *** The monitor lady smiled very nicely and tousled his hair and said, "Andrew, I suppose by now you're just absolutely sick of having that horrid monitor. Well, I have good news for you. That monitor is going to come out today. We're going to just take it right out, and it won't hurt a bit." Ender nodded. It was a lie, of course, that it wouldn't hurt a bit. But since adults always said it when it was going to hurt, he could count on that statement as an accurate prediction of the future. Sometimes lies were more dependable than the truth. "So if you'll just come over here, Andrew, just sit right up here on the examining table.
    [Show full text]