Login

Publications  •  Project Statistics

Glossary  •  Schools  •  Disciplines
People Search: 
   
Title/Abstract Search: 

Dissertation Information for Filippo Menczer

NAME:
- Filippo Menczer

DEGREE:
- Ph.D.

DISCIPLINE:
- Computer Science

SCHOOL:
- University of California, San Diego (USA) (1998)

ADVISORS:
- None

COMMITTEE MEMBERS:
- None

MPACT Status: Incomplete - Not_Inspected

Title: Life-like agents: Internalizing local cues for reinforcement learning and evolution

Abstract: "As computing and information environments become more complex, there is a need for tools that assist users in ways that scale with the growth of such environments and adapt to both the personal preferences of the user and the changes in user and environmental conditions. In this thesis I attempt to abstract some core principles from ecological adaptive processes, and cast them into a machine learning framework. The thesis focuses on two mechanisms--local selection and internalization--by which evolutionary algorithms can be: (i) efficiently parallelized, for adaptation in distributed environmental and (ii) integrated with both unsupervised and supervised learning, for adaptation at different spatial and temporal scales in dynamic and heterogeneous environments.

These methods have been applied in the construction of populations of adaptive information agents (InfoSpiders). Agents search online for information relevant, to the user, by traversing links in distributed information environments such as the Web. I demonstrate feasibility of the approach by comparing the performance of InfoSpiders with both exhaustive search and other heuristics, on well-defined hypertext environments. InfoSpiders are capable of capturing essential features of their spatially and temporally local context, thus adapting to complex dynamic environments and user requests. InfoSpiders complement search engines with respect to several of their difficulties. Engines provide global starting points, based on statistical features of the search space (words): InfoSpiders use topological features (links) to guide their subsequent search.

The same principles have also been applied to build better models of ecological adaptation. I have developed a model and simulator called LEE (Latent Energy Environments), whereby agents adapt by a genetic algorithm based on local selection. This simple individual-based model results in the emergence of density-dependent evolution, and allows the simultaneous study of adaptation at the individual and population level. LEE has been used to analyze the roles of different types of environmental structure in shaping the selective pressures that allow complex collective behaviors to evolve. I also report on the results of other experiments, aimed at observing the interactions between different evolving traits--behaviors, morphology. and maturation and different forms of individual plasticity--prediction, reinforcement learning, and imitation."

MPACT Scores for Filippo Menczer

A = 0
C = 2
A+C = 2
T = 0
G = 0
W = 0
TD = 0
TA = 0
calculated 2010-12-13 19:01:02

Advisors and Advisees Graph