Differences

This shows you the differences between two versions of the page.

--- teaching:mfe:ia [2013/03/18 10:02]
bersini
+++ teaching:mfe:ia [2013/04/03 11:06]
stuetzle
@@ Line 1: / Line 1: @@
-====== MFE 2012-2013 : Intelligence Artificielle ======
+====== MFE 2013-2014 : Intelligence Artificielle ======
 ===== Introduction =====
@@ Line 216: / Line 216: @@
-===== Stochastic local search algorithms for weighted maximum clique problems. ======
-The Maximum Clique Problem is an NP-hard combinatorial optimisation problem that asks to find the biggest completely
+===== Stochastic Local Search heuristics for solving NP-complete puzzles. ======
-connected component of a graph. It has relevant applications in information retrieval, computer vision, social network
-analysis, computational biochemistry, bioinformatics and genomics.
-Among the possible generalisations of the problem there is the Vertex Weighted and Edge Weighted Maximum Clique which asks to find the clique of maximum weight. Being generalisations they are also NP-hard. The goal of the project is to devise heuristic algorithms or adapt existing algorithms of the Maximum Clique for weighted version.
+This project is about single player games (puzzles) and the design of algorithms for tackling hard combinatorial optimisation problems.
+Example puzzles are: [[http://en.wikipedia.org/wiki/Light_Up|Light Up]], [[http://en.wikipedia.org/wiki/Mastermind_(board_game)|Mastermind]], [[http://en.wikipedia.org/wiki/Minesweeper_(video_game)|Minesweeper]], etc.
+The student will learn how to design and implement a Stochastic Local Search algorithm to solve NP-complete puzzles. The student will also learn how to analyse the performaces of the algorithm and perform statistically sound comparisons with the other algorithms available in literature.
 Required skills: good knowledge of C or C++ programming.
@@ Line 232: / Line 232: @@
+===== Analysis of Local Optima Networks for the Max-Clique problem. ======
-===== Stochastic Local Search heuristics for solving NP-complete puzzles. ======
+Stochastic Local Search algorithms search for optimal solutions in a large space of candidate solutions. Local Optima Networks (LON) are representations of search landscapes of combinatorial optimisation problems. In these networks, nodes are local optima of the problem, and edges are weighted transitions between the optima. These networks model in a more compact way the properties of the larger search spaces they represent.
-This project is about single player games (puzzles) and the design of algorithms for tackling hard combinatorial optimisation problems.
+The goal of this project is to build LON of small instances of the Maximum Clique (MC) problem, and measure properties that could illustrate the differences between instance families.
-Example puzzles are: [[http://en.wikipedia.org/wiki/Light_Up|Light Up]], [[http://en.wikipedia.org/wiki/Mastermind_(board_game)|Mastermind]], [[http://en.wikipedia.org/wiki/Minesweeper_(video_game)|Minesweeper]], etc.
+The MC problem is an NP-hard combinatorial optimisation problem that asks to find the biggest completely connected component of a graph. It has relevant applications in information retrieval, computer vision, social network analysis, computational biochemistry, bioinformatics and genomics.
-The student will learn how to design and implement a Stochastic Local Search algorithm to solve NP-complete puzzles. The student will also learn how to analyse the performaces of the algorithm and perform statistically sound comparisons with the other algorithms available in literature.
 Required skills: good knowledge of C or C++ programming.
@@ Line 248: / Line 249: @@
-===== Analysis of Local Optima Networks for the Max-Clique problem. ======
+===== Feature Extraction and Automatic Algorithm Selection. ======
-Stochastic Local Search algorithms search for optimal solutions in a large space of candidate solutions. Local Optima Networks (LON) are representations of search landscapes of combinatorial optimisation problems. In these networks, nodes are local optima of the problem, and edges are weighted transitions between the optima. These networks model in a more compact way the properties of the larger search spaces they represent.
+The performance of (Stochastic Local Search) algorithms for a given problem depends on the algorithm design and on the setting of the algorithm's parameter. Given a heterogeneous set of instances for a given problem a good algorithm design (or parameter configuration) for one instance is not necessary the best design for all instances. On the contrary a tuning of an algorithm on a specific family of similar instances may affect negatively its performance on other families of instances.
-The goal of this project is to build LON of small instances of the Maximum Clique (MC) problem, and measure properties that could illustrate the differences between instance families.
+The thesis will focus on devising automatic methods for extracting features from the instances, select the relevant features, and learning (in the framework of multi-class classification) the
+relationship, if there is one, between the instances features and the best algorithm for the instance. The results will be instrumental for algorithm selection or the creation of portfolios of complementary algorithms suitable for large sets of diverse instances for a given problem.
-The MC problem is an NP-hard combinatorial optimisation problem that asks to find the biggest completely connected component of a graph. It has relevant applications in information retrieval, computer vision, social network analysis, computational biochemistry, bioinformatics and genomics.
+Required skills: good knowledge of C or C++ programming and of a scripting language (e.g., python); good knowledge of machine learning methods would also be helpful.
-Required skills: good knowledge of C or C++ programming.
@@ Line 263: / Line 263: @@
     * [[http://iridia.ulb.ac.be/~stuetzle|Thomas Stützle (IRIDIA)]]
     * [[http://iridia.ulb.ac.be/~fmascia|Franco Mascia (IRIDIA)]]