Research.HomePage History

Hide minor edits - Show changes to markup

December 16, 2013, at 04:18 PM by 131.175.28.194 -
Added lines 3-23:

Genetics-Based Machine Learning

Genetics-Based Machine Learning (GBML) is a machine learning paradigm introduced by Holland in 1976 based on evolutionary computation. In this paradigm, the learning is viewed as a process of ongoing adaptation to an unknown environment which provides feedback in terms of numerical reward. The incoming reward is then used to guide the evolution of a population of condition-action-prediction rules, called classifiers, which represents the solution to the target problem. Each classifier represents a small piece of the overall solution: the condition identifies a problem subspace; the action represents a decision to take in the problem subspace identified by the classifier condition; the prediction estimates how valuable the classifier is in terms of problem solution. My research in this area focused mainly on the following topics:

  • theoretical analysis of the GBML systems
  • design and extension of the classifier prediction model
  • adapting the classifier prediction model to the problem
  • GBML systems applied to the design space exploration of embedded systems
  • implementation of GBML systems on GPUs

Computational Intelligence and Games

The Electronic Entertainment industry grew very fast and attracted a lot of researchers in the recent years. In this area, my research interests are articulated in two main directions: video games as testbed for Computa- tional Intelligence (CI) methods and the automatic game content generation.

Video Game as Testbed for CI

Modern video games are at the same time a fascinating application domain and an ideal testbed for the CI methods. My main contribution in this area is the design and the organization of the Simulated Car Racing Competition, a scientific competition where the goal is developing (by means of a CI approach) a controller for The Open Racing Car Simulator (TORCS), an open-source racing game. So far, the Simulated Car Racing has been used as research platform in approximately 20 published works (in proceedings of international conferences as well in international journals) in the game research community.

Automatic Game Content Generation

During the development of a modern game a major part of the avail- able resources is used to create the game content, such as the game mechanics, the environments and the characters. In order to develop ground-breaking new games, the industry is in need of reliable and effective tools for creating contents capable of engaging the customers. Moreover, the broadening of the customer base poses new additional challenges to the game industry and demands and for individualization to the abilities and needs of the single customer. In this scenario, my research interests involve the application of CI methods (i) to develop characters at the same time challenging and believable, (ii) to enable learning and adaptivity in games, and (iii) to generate game content, that is both innovative and entertaining.

(:if false:) My research interests are in the area of Machine Learning and Evolutionary computation.

Changed lines 71-72 from:
  • machine learning for modeling and estimating the cost and the performance of embedded systems
to:
  • machine learning for modeling and estimating the cost and the performance of embedded systems

(:ifend:)

October 04, 2008, at 02:34 PM by 79.11.1.50 -
Changed line 47 from:

My research focus mainly on:

to:

My research focus is mainly on:

February 15, 2008, at 02:18 PM by 131.175.124.43 -
Changed line 25 from:

Machine learning has many interesting application in modern computer gaims.

to:

Machine learning has many interesting application in modern computer games.

February 15, 2008, at 11:08 AM by 131.175.124.43 -
Changed line 26 from:

The research aims to investigate the following research directions:

to:

The research aims to investigate the following issues:

February 15, 2008, at 10:25 AM by 131.175.124.43 -
February 15, 2008, at 10:25 AM by 131.175.124.43 -
Changed lines 31-32 from:

More details will be available on the | CIG@PoliMI web site.

to:

More details will be available on the CIG@PoliMI web site.

February 15, 2008, at 10:24 AM by 131.175.124.43 -
Changed line 26 from:

In particular my research focus on:

to:

The research aims to investigate the following research directions:

Added lines 31-32:

More details will be available on the | CIG@PoliMI web site.

Changed line 35 from:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain.

to:

I am interested in the application of machine learning techniques for solving relevant problems in the financial domain.

Changed line 37 from:

Both these problems can be modeled as reinforcement learning problems and

to:

Both these problems have been modeled as reinforcement learning problems and

Changed line 47 from:

My research focus mainly on three directions:

to:

My research focus mainly on:

February 13, 2008, at 03:40 PM by 131.175.124.43 -
Added lines 21-30:

Machine Learning for Modern Computer Games

Machine learning has many interesting application in modern computer gaims. In particular my research focus on:

  • exploiting machine learning for modeling human players
  • applying machine learning for developing interesting behaviors of non player characters
  • using machine learning to adapt games to user preference and to improve game experience
Changed lines 32-35 from:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain.
Two interesting examples are the portfolio optimization and the the one-way trading.
Both these problems can be modeled as reinforcement learning problems and \\

to:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain. Two interesting examples are the portfolio optimization and the the one-way trading. Both these problems can be modeled as reinforcement learning problems and

Deleted lines 38-45:

Machine Learning for Modern Computer Games

Machine learning has many interesting application in modern computer gaims. In particular my research focus on:
(i) exploiting machine learning for modeling human players
(ii) applying machine learning for developing interesting behaviors of non player characters
(iii) using machine learning to adapt games to user preference and to improve game experience

February 13, 2008, at 03:39 PM by 131.175.124.43 -
Changed line 37 from:

Machine Learning and Evolutionary Computation for Embedded System Design

to:

ML and EC for Embedded Systems Design

February 13, 2008, at 03:37 PM by 131.175.124.43 -
February 13, 2008, at 03:37 PM by 131.175.124.43 -
Changed line 33 from:

(i) exploiting machine learning for modeling human players

to:

(i) exploiting machine learning for modeling human players\\

February 13, 2008, at 03:37 PM by 131.175.124.43 -
February 13, 2008, at 03:37 PM by 131.175.124.43 -
Changed lines 33-36 from:
(i) exploiting machine learning for modeling human players
(ii) applying machine learning for developing interesting behaviors of non player characters
-> (iii) using machine learning to adapt games to user preference and to improve game experience
to:

(i) exploiting machine learning for modeling human players (ii) applying machine learning for developing interesting behaviors of non player characters
(iii) using machine learning to adapt games to user preference and to improve game experience

February 13, 2008, at 03:36 PM by 131.175.124.43 -
Changed lines 32-36 from:

In particular my research focus on:

  • exploiting machine learning for modeling human players
  • applying machine learning for developing interesting behaviors of non player characters
  • using machine learning to adapt games to user preference and to improve game experience
to:

In particular my research focus on:
-> (i) exploiting machine learning for modeling human players

(ii) applying machine learning for developing interesting behaviors of non player characters
-> (iii) using machine learning to adapt games to user preference and to improve game experience
February 13, 2008, at 03:35 PM by 131.175.124.43 -
Changed line 30 from:
to:
February 13, 2008, at 03:32 PM by 131.175.124.43 -
Changed lines 22-27 from:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain. Two interesting examples are the portfolio optimization and the the one-way trading. Both these problems can be modeled as reinforcement learning problems and therefore they can be solved either with usual reinforcement learning or with

to:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain.
Two interesting examples are the portfolio optimization and the the one-way trading.
Both these problems can be modeled as reinforcement learning problems and
therefore they can be solved either with usual reinforcement learning or with

Added line 30:
Added line 38:
February 13, 2008, at 03:26 PM by 131.175.124.43 -
Changed line 22 from:
to:
February 13, 2008, at 03:21 PM by 131.175.124.43 -
Changed line 22 from:
to:
February 13, 2008, at 03:21 PM by 131.175.124.43 -
Changed line 22 from:
to:
February 13, 2008, at 03:20 PM by 131.175.124.43 -
Changed line 22 from:

[%rfloat]

to:
February 13, 2008, at 03:14 PM by 131.175.124.43 -
Added line 22:

[%rfloat]

February 12, 2008, at 11:10 PM by 213.230.129.20 -
Added line 4:
Changed lines 7-9 from:

In learning classifier systems,

  the learning is viewed a process of ongoing adaptation
  to an unknown environment which provides feedback in terms of numerical reward.
to:

In learning classifier systems, the learning is viewed a process of ongoing adaptation to an unknown environment which provides feedback in terms of numerical reward.

Changed lines 10-11 from:
  guide the evolution of a population of condition-action-prediction
  rules, called classifiers, which represents the solution to the target problem.
to:

guide the evolution of a population of condition-action-prediction rules, called classifiers, which represents the solution to the target problem.

Changed lines 12-15 from:
  the condition identifies a problem subspace;
  the action represents a decision to take in the problem subspace
  identified by the classifier condition;
  the prediction estimates how valuable the classifier is in terms of problem solution.
to:

the condition identifies a problem subspace; the action represents a decision to take in the problem subspace identified by the classifier condition; the prediction estimates how valuable the classifier is in terms of problem solution.

Changed lines 24-25 from:

Two interesting examples are the portfolio optimization and the the

  one-way trading.
to:

Two interesting examples are the portfolio optimization and the the one-way trading.

February 12, 2008, at 11:09 PM by 213.230.129.20 -
Changed lines 32-48 from:

Nell'ambito dei videogiochi, la ricerca ha diversi obiettivi: (i) l'utilizzo di tecniche di apprendimento automatico per modellizzare il comportamento di giocatori umani; (ii) la comparazione di tecniche di apprendimento per rinforzo e di computazione evolutiva per sviluppare l'intelligenza artificiale nei videogiochi; (iii) l'utilizzo di tecniche di apprendimento automatico sia per adattare il gioco alle preferenze dell'utente

  sia per sviluppare moduli in grado di facilitare l'esperienza di gioco.

\subsection{Computazione Evolutiva ed Apprendimento Automatico per la Progettazione di Sistemi Dedicati} La crescente diffusione dei sistemi dedicati e l'aumentare della loro complessit\`a rende sempre pi\`u importante

  lo sviluppo di strumenti di supporto per lo sviluppo di questi sistemi.

In generale, la progettazione dei sistemi dedicati comporta la risoluzione di diversi problemi di ottimizzazione

  (ad esempio l'allocazione di risorse e lo scheduling di operazioni) che sono fra di loro fortemente 
  dipendenti e difficili da modellizzare in maniera analitica.

In questo scenario le tecniche di computazione evolutiva e di apprendimento automatico possono essere applicate per risolvere alcuni di questi problemi in maniera efficace. Le direzioni di ricerca sviluppate sono principalmente tre: (i) l'applicazione di algoritmi genetici multi-obiettivo per l'esplorazione dello spazio di progetto in presenza di obiettivi in contrasto fra loro; (ii) l'utilizzo di algoritmi genetici avanzati per l'ottimizzazione dell'allocazione delle risorse in sistemi dedicati eterogenei e riconfigurabili; (iii) l'applicazione di tecniche di apprendimento automatico per la modellizzazione e la stima di costi e prestazioni.

to:

Machine learning has many interesting application in modern computer gaims. In particular my research focus on:

  • exploiting machine learning for modeling human players
  • applying machine learning for developing interesting behaviors of non player characters
  • using machine learning to adapt games to user preference and to improve game experience

Machine Learning and Evolutionary Computation for Embedded System Design

Today embedded system are widely used and more and more complex. This turns out in an increasing need of automated tool for the design of such systems. In general, the design of an embedded system involves the solution of several difficult and highly interdependent problems. In this scenario evolutionary computation and machine learning can be applied very effectively. My research focus mainly on three directions:

  • multi-objective genetic algorithms for the design space exploration to deal with conflicting objectives
  • probabilistic genetic algorithms for the resource allocation in the design of heterogeneous and reconfigurable embedded systems
  • machine learning for modeling and estimating the cost and the performance of embedded systems
February 12, 2008, at 10:50 PM by 213.230.129.20 -
Changed lines 23-31 from:
 La ricerca si propone di studiare l'applicazione di tecniche di apprendimento automatico, tra cui i sistemi a classificatori, per risolvere problemi nell'ambito dei mercati finanziari e dei videogiochi commerciali. 

In particolare, nell'ambito dei mercati finanziari la ricerca si sviluppa su due problemi

  specifici: 

l'ottimizzazione di un portafoglio di titoli azionari e il problema della \emph{Trade Execution}

  (noto anche come \emph{One-Way Trading}).

Entrambi i problemi sono stati modellizzati come problemi di apprendimento per rinforzo. La ricerca si propone di confrontare tecniche di apprendimento per rinforzo e tecniche basate sulla computazione evolutiva per risolvere questi problemi.

to:

We are interested in the application of machine learning techniques for solving relevant problems in the financial domain. Two interesting examples are the portfolio optimization and the the

  one-way trading.

Both these problems can be modeled as reinforcement learning problems and therefore they can be solved either with usual reinforcement learning or with genetics-based machine learning techniques, like learning classifier systems.

February 12, 2008, at 10:43 PM by 213.230.129.20 -
Changed lines 1-19 from:

'attuale attivit\`a di ricerca si articola principalmente nelle seguenti aree.

\subsection{Apprendimento Automatico con Tecniche di Computazione Evolutiva} \`E il tema della tesi di dottorato e ha riguardato lo studio dei sistemi a classificatori, un paradigma di apprendimento automatico in cui la conoscenza \`e rappresentata da un insieme di regole. In particolare la ricerca si \`e focalizzata sull'utilizzo dei sistemi a classificatori per l'addestramento di agenti autonomi in problemi di apprendimento per rinforzo. La ricerca \`e stata sviluppata principalmente nelle seguenti direzioni: (i) analisi delle capacit\`a di generalizzazione dei sistemi studiati; % (iii) utilizzo del filtro di Kalman e metodi ai minimi quadrati per stimare i parametri e l'errore regole evolute nei sistemi a classificatori; (ii) formalizzazione dei sistemi a classificatori come paradigma di apprendimento per rinforzo; (iii) estensione dei sistemi a classificatori con tecniche di apprendimento per rinforzo specifiche per problemi complessi; (iv) estensione dei sistemi a classificatori a problemi di classificazione supervisionata; (v) utilizzo dei sistemi a classificatori per evolvere ensemble di Reti Neurali o Support Vector Machines

  per risolvere problemi di apprendimento per rinforzo, di regressione e di classificazione.

\subsection{Apprendimento Automatico nei Videogiochi e nei Mercati Finanziari} \noindent La ricerca si propone di studiare l'applicazione di tecniche di apprendimento automatico, tra cui i sistemi a classificatori, per risolvere problemi nell'ambito dei mercati finanziari e dei videogiochi commerciali.

to:

My research interests are in the area of Machine Learning and Evolutionary computation.

Learning Classifier Systems (Genetics-Based Machine Learning)

Learning classifier systems are a machine learning paradigm introduced by Holland in 1976 based on evolutionary computation. In learning classifier systems,

  the learning is viewed a process of ongoing adaptation
  to an unknown environment which provides feedback in terms of numerical reward.

Learning classifier systems use the incoming reward to

  guide the evolution of a population of condition-action-prediction
  rules, called classifiers, which represents the solution to the target problem.

Each classifier represents a small piece of the overall solution:

  the condition identifies a problem subspace;
  the action represents a decision to take in the problem subspace
  identified by the classifier condition;
  the prediction estimates how valuable the classifier is in terms of problem solution.

In particular, my research focused on classifier prediction models:

  • how prediction model affect the overall performance?
  • which prediction model should be used (e.g., neural networks or support vector machines) ?
  • how to exploit the genetic algorithms to adapt the prediction model to the problem ?

Machine Learning in Finance

 La ricerca si propone di studiare l'applicazione di tecniche di apprendimento automatico, tra cui i sistemi a classificatori, per risolvere problemi nell'ambito dei mercati finanziari e dei videogiochi commerciali. 
Added line 32:

Machine Learning for Modern Computer Games

February 12, 2008, at 09:01 PM by 213.230.129.20 -
Added lines 1-43:

'attuale attivit\`a di ricerca si articola principalmente nelle seguenti aree.

\subsection{Apprendimento Automatico con Tecniche di Computazione Evolutiva} \`E il tema della tesi di dottorato e ha riguardato lo studio dei sistemi a classificatori, un paradigma di apprendimento automatico in cui la conoscenza \`e rappresentata da un insieme di regole. In particolare la ricerca si \`e focalizzata sull'utilizzo dei sistemi a classificatori per l'addestramento di agenti autonomi in problemi di apprendimento per rinforzo. La ricerca \`e stata sviluppata principalmente nelle seguenti direzioni: (i) analisi delle capacit\`a di generalizzazione dei sistemi studiati; % (iii) utilizzo del filtro di Kalman e metodi ai minimi quadrati per stimare i parametri e l'errore regole evolute nei sistemi a classificatori; (ii) formalizzazione dei sistemi a classificatori come paradigma di apprendimento per rinforzo; (iii) estensione dei sistemi a classificatori con tecniche di apprendimento per rinforzo specifiche per problemi complessi; (iv) estensione dei sistemi a classificatori a problemi di classificazione supervisionata; (v) utilizzo dei sistemi a classificatori per evolvere ensemble di Reti Neurali o Support Vector Machines

  per risolvere problemi di apprendimento per rinforzo, di regressione e di classificazione.

\subsection{Apprendimento Automatico nei Videogiochi e nei Mercati Finanziari} \noindent La ricerca si propone di studiare l'applicazione di tecniche di apprendimento automatico, tra cui i sistemi a classificatori, per risolvere problemi nell'ambito dei mercati finanziari e dei videogiochi commerciali.

In particolare, nell'ambito dei mercati finanziari la ricerca si sviluppa su due problemi

  specifici: 

l'ottimizzazione di un portafoglio di titoli azionari e il problema della \emph{Trade Execution}

  (noto anche come \emph{One-Way Trading}).

Entrambi i problemi sono stati modellizzati come problemi di apprendimento per rinforzo. La ricerca si propone di confrontare tecniche di apprendimento per rinforzo e tecniche basate sulla computazione evolutiva per risolvere questi problemi.

Nell'ambito dei videogiochi, la ricerca ha diversi obiettivi: (i) l'utilizzo di tecniche di apprendimento automatico per modellizzare il comportamento di giocatori umani; (ii) la comparazione di tecniche di apprendimento per rinforzo e di computazione evolutiva per sviluppare l'intelligenza artificiale nei videogiochi; (iii) l'utilizzo di tecniche di apprendimento automatico sia per adattare il gioco alle preferenze dell'utente

  sia per sviluppare moduli in grado di facilitare l'esperienza di gioco.

\subsection{Computazione Evolutiva ed Apprendimento Automatico per la Progettazione di Sistemi Dedicati} La crescente diffusione dei sistemi dedicati e l'aumentare della loro complessit\`a rende sempre pi\`u importante

  lo sviluppo di strumenti di supporto per lo sviluppo di questi sistemi.

In generale, la progettazione dei sistemi dedicati comporta la risoluzione di diversi problemi di ottimizzazione

  (ad esempio l'allocazione di risorse e lo scheduling di operazioni) che sono fra di loro fortemente 
  dipendenti e difficili da modellizzare in maniera analitica.

In questo scenario le tecniche di computazione evolutiva e di apprendimento automatico possono essere applicate per risolvere alcuni di questi problemi in maniera efficace. Le direzioni di ricerca sviluppate sono principalmente tre: (i) l'applicazione di algoritmi genetici multi-obiettivo per l'esplorazione dello spazio di progetto in presenza di obiettivi in contrasto fra loro; (ii) l'utilizzo di algoritmi genetici avanzati per l'ottimizzazione dell'allocazione delle risorse in sistemi dedicati eterogenei e riconfigurabili; (iii) l'applicazione di tecniche di apprendimento automatico per la modellizzazione e la stima di costi e prestazioni.