Optimization of sequences of observations for the automatic diagnostic

For the final project of the second year of the master, I worked in a project based in the "Decision Theoretic Troubleshooting".

It consists in a interactive application with an user interface, when it is represented a system with its components. For each component we have its probability of running, and all the system is represented by a Bayesian network.

The application help the user to generate a diagnostic for the reparation of the system finding the states with anomalies and his dependence and repercussion among the others components.

The entire application was developed in python, with the PyQt and gnuplot libraries for the user interface and the PyAgrum library for the representation of the Bayesian network.

Paris - France, January - March 2013

Ontologies Mining

As a personal project I worked in the development of a system which search separates articles from a web site. The system takes the most important words which describe a concept amongst different writers and automatically the system builds an ontology (RDF graph) with all of the concepts with the sentences and the articles where they were found.

For the system’s development I used the python programming language using for the selection of the most important words which describe a concept, the Bag of words (BoW) algorithm where I constructed a histogram with the words used in the different articles and its repetition numbers. Each histogram element it was clustered using the k-means algorithm amongst “so repeated”, “normally repeated”, and “not repeated”, filtering and only taking the words classed as “normally repeated”.

Paris - France, February 2013

Recommendation system

As an academic project I worked in the development using the python programming language of a recommendation system which has as an entry the preferences and class of several users in a web site.

I used the naive Bayes algorithm when we suppose all the variables independent, with the maximum likelihood and the priori knowledge approaches to determinate the probabilities of belonging for each class in such a way that the system could to predict the class of a new user who doesn’t have all the preferences and furthermore to predict the preferences in absence.

I coded as well an approach using the tree-augmented naive model (TAN) algorithm building a Bayesian network which we learned the mutual information between the variables to predict the class which a user belongs.

Paris - France, January 2013

System for the analysis of web traces and clustering using the k-means algorithm

As an academic project I worked in the development of a system in the Java programming language with the Swing library for the user interface, which has as an entry a log document type “Combined log” where we take for each request the user id. We used an interval of 30 minutes to set a session. It means that several request with the same user in an interval between them lower than 30 minutes compose a session.

For the clustering of the different sessions I used the k-means algorithm with the numbers of clusters and the kind of distance as parameters. For the different kind of distances I coded the Euclidean distance, the cosine measure, and the Jaccard distance for the calculation at the moment to compare the sessions.

So that I clustered the sessions in different groups having common requests in such a way that we could to determinate statistics such as: the sites with the lowest and highest concurrence, predictions about links for the users, and relations between links.

Paris - France, December 2012

Approach to robotics using reinforcement learning

As an academic project I coded 3 reinforcement learning algorithms to learn how to a robot could to walk.

I used the V function with the Bellman equation, the Q function with a reformulation of the Bellman equation, and the Q learning algorithm with an approach E-greedy.

For the implementation of the algorithms I used a reward vector with a punishment when the robot goes down and goes back and with a reward when the robot goes forward. Likewise I used a transition vector with the different possible robot actions having false for the transitions which make the robot falls over.

Paris - France, November 2012

Multi-agents video-game using algorithms in reinforcement learning

As a personnel project I worked in the development of a video-game of about several agents who search in a laboratory for different components with the aim of create a nuclear bomb. The user player has to stop them to save the world.

For the development of the video-game I used the programming language Java with the swing library for the user interface and the JADE library for behavior programming in multi-agents.

For the artificial intelligence in agents I coded the MDP (Markov decision process) algorithm which allows to each agent how to find the shorter trail to the nearer bomb component, synchronizing and distributing the tasks for each agent using the Zeuthon algorithm.

For the MDP algorithm I used a reward’s vector with punishment for the position of the user player and rewards for the positions of the bomb components and an action's vector with the possible actions for the current agent.

Paris - France, October 2012

Multi-agent simulation platform for modeling agent’s behaviors in organizations

As an academic project I developed a multi-agent platform using netlogo script programming to simulate and modeling the effort and profit exerted by heterogeneous agents in an organization.

The platform consists in a user interface with the parameters as follows:

-              -10 sets of agents, each on with a different behavior, we can choose how many agents for each type of            agent:

o   null effort: this agent always exerts the same almost null effort

o   shrinking effort: this agent halves the effort provided by its last partner

o   replicator: this agent exerts the same effort its last partner exerted in the previous interaction

o   rational: this agent exerts the best reply for its last partner effort

o   profit comparator: this agent compares its profit to its last partner's one; it increases its effort if it gave a higher profit

o   high effort: this agent always exerts the same high effort

o   average rational: this agent exerts the best reply to the average effort of its partners

o   winner imitator: this agent starts with high effort but copies its partner's effort when this one proves to yield a higher profit

o   effort comparator: this agent compares its effort to its last partner's one; it increases its effort if it is inferior to its partner's one and vice versa

o   averager: it averages its effort with its last partner's effort

-              -A noise percentage at the moment of the communication between the agents.

The system reaches to find the Nash equilibrium in the society, in a way that each agent maximize his effort without minimize his profit.

Paris - France, September 2012