Co-evolutionary and Reinforcement Learning Techniques Applied to Computer Go players

The objective of this thesis is model some processes from the nature as evolution and co-evolution, and proposing some techniques that can ensure that these learning process really happens and useful to solve some complex problems as Go game. The Go game is ancient and very complex game with simple rules which still is a challenge for the Artificial Intelligence. This dissertation cover some approaches that were applied to solve this problem, proposing solve this problem using competitive and cooperative co-evolutionary learning methods and other techniques proposed by the author. To study, implement and prove these methods were used some neural networks structures, a framework free available and coded many programs. The techniques proposed were coded by the author, performed many experiments to find the best configuration to ensure that co-evolution is progressing and discussed the results. Using co-evolutionary learning processes can be observed some pathologies which could impact co-evolution progress. In this dissertation is introduced some techniques to solve pathologies as loss of gradients, cycling dynamics and forgetting. According to some authors, one solution to solve these co-evolution pathologies is introduce more diversity in populations that are evolving. In this thesis is proposed some techniques to introduce more diversity and some diversity measurements for neural networks structures to monitor diversity during co-evolution. The genotype diversity evolved were analyzed in terms of its impact to global fitness of the strategies evolved and their generalization. Additionally, it was introduced a memory mechanism in the network neural structures to reinforce some strategies in the genes of the neurons evolved with the intention that some good strategies learned are not forgotten. In this dissertation is presented some works from other authors in which cooperative and competitive co-evolution has been applied. The Go board size used in this thesis was 9x9, but can be easily escalated to more bigger boards.The author believe that programs coded and techniques introduced in this dissertation can be used for other domains.

Approximate Relational Reasoning for Probabilistic Programs

La seguridad verificada es una metodología para demostrar propiedades de seguridad de los sistemas informáticos que se destaca por las altas garantías de corrección que provee. Los sistemas informáticos se modelan como programas probabilísticos y para probar que verifican una determinada propiedad de seguridad se utilizan técnicas rigurosas basadas en modelos matemáticos de los programas. En particular, la seguridad verificada promueve el uso de demostradores de teoremas interactivos o automáticos para construir demostraciones completamente formales cuya corrección es certificada mecánicamente (por ordenador). La seguridad verificada demostró ser una técnica muy efectiva para razonar sobre diversas nociones de seguridad en el área de criptografía. Sin embargo, no ha podido cubrir un importante conjunto de nociones de seguridad “aproximada”. La característica distintiva de estas nociones de seguridad es que se expresan como una condición de “similitud” entre las distribuciones de salida de dos programas probabilísticos y esta similitud se cuantifica usando alguna noción de distancia entre distribuciones de probabilidad. Este conjunto incluye destacadas nociones de seguridad de diversas áreas como la minería de datos privados, el análisis de flujo de información y la criptografía. Ejemplos representativos de estas nociones de seguridad son la indiferenciabilidad, que permite reemplazar un componente idealizado de un sistema por una implementación concreta (sin alterar significativamente sus propiedades de seguridad), o la privacidad diferencial, una noción de privacidad que ha recibido mucha atención en los últimos años y tiene como objetivo evitar la publicación datos confidenciales en la minería de datos. La falta de técnicas rigurosas que permitan verificar formalmente este tipo de propiedades constituye un notable problema abierto que tiene que ser abordado. En esta tesis introducimos varias lógicas de programa quantitativas para razonar sobre esta clase de propiedades de seguridad. Nuestra principal contribución teórica es una versión quantitativa de una lógica de Hoare relacional para programas probabilísticos. Las pruebas de correción de estas lógicas son completamente formalizadas en el asistente de pruebas Coq. Desarrollamos, además, una herramienta para razonar sobre propiedades de programas a través de estas lógicas extendiendo CertiCrypt, un framework para verificar pruebas de criptografía en Coq. Confirmamos la efectividad y aplicabilidad de nuestra metodología construyendo pruebas certificadas por ordendor de varios sistemas cuyo análisis estaba fuera del alcance de la seguridad verificada. Esto incluye, entre otros, una meta-construcción para diseñar funciones de hash “seguras” sobre curvas elípticas y algoritmos diferencialmente privados para varios problemas de optimización combinatoria de la literatura reciente. ABSTRACT The verified security methodology is an emerging approach to build high assurance proofs about security properties of computer systems. Computer systems are modeled as probabilistic programs and one relies on rigorous program semantics techniques to prove that they comply with a given security goal. In particular, it advocates the use of interactive theorem provers or automated provers to build fully formal machine-checked versions of these security proofs. The verified security methodology has proved successful in modeling and reasoning about several standard security notions in the area of cryptography. However, it has fallen short of covering an important class of approximate, quantitative security notions. The distinguishing characteristic of this class of security notions is that they are stated as a “similarity” condition between the output distributions of two probabilistic programs, and this similarity is quantified using some notion of distance between probability distributions. This class comprises prominent security notions from multiple areas such as private data analysis, information flow analysis and cryptography. These include, for instance, indifferentiability, which enables securely replacing an idealized component of system with a concrete implementation, and differential privacy, a notion of privacy-preserving data mining that has received a great deal of attention in the last few years. The lack of rigorous techniques for verifying these properties is thus an important problem that needs to be addressed. In this dissertation we introduce several quantitative program logics to reason about this class of security notions. Our main theoretical contribution is, in particular, a quantitative variant of a full-fledged relational Hoare logic for probabilistic programs. The soundness of these logics is fully formalized in the Coq proof-assistant and tool support is also available through an extension of CertiCrypt, a framework to verify cryptographic proofs in Coq. We validate the applicability of our approach by building fully machine-checked proofs for several systems that were out of the reach of the verified security methodology. These comprise, among others, a construction to build “safe” hash functions into elliptic curves and differentially private algorithms for several combinatorial optimization problems from the recent literature.