Back
to indexFilippo Menczer and Richard K. Belew
Summary
This paper is provides an extensive analysis of adaptive search agents, providing details of two implementations: local selection for adaptation and internalization of learned environmental signals. The utility of linkage topology is explored, and the use of neighborhoods of relevant pages is incorporated into the selection of which link an agent should follow. Agents may internalize the environmental structure through reinforcement learning of which pages provide more energy, or are hit pages. Two measures, R, the correlation of good nodes, and G, the density of good nodes, are determined to be critical factors in search performance. The interesting step in the detailed InfoSpiders algorithm includes the use of Q-learning, which adjusts the weights of the keywords used as input for the neural net. Local selection is used by InfoSpiders to eliminate unproductive agents and reward productive agents. The efficiency of this approach is experimentally confirmed.
This paper provides sufficient detail to allow a reader to run out and code up a reasonable approximation of the InfoSpiders program. The experiments are also reproducible, given an instance of the EB. Overall, this paper provides a solid foundation, with sufficient overview of the potential issues that may occur when building an adaptive distributed agent.
Keywords
InfoSpiders, distributed information retrieval, evolutionary algorithms, local selection, internalization, reinforcement learning, Q-learning, neural networks, relevance feedback, linkage topology, relevance autocorrelation, scalability, context, personalization, localization, selective query expansion, dimensionality reduction, graph search, browsing, on-line, mobile, situated, adaptive agents
Methods
Linkage topology is explored by browsing agents following links based on which will most likely lead to a relevant page, given the current page is relevant, which is defined to be the relevance autocorrelation. These agents are extrememly scalable, due to their distributed, parallelizable nature. Local selection is variable based on the carrying capacity of the environment. Agents are able to internalize the state of the environment based on reinforcement learning, then Q-learning the weights of the neural nets. Agents generate offspring through a combination of cloning, mutation, and crossover, based on the environmental characteristics.
Rating
8
Bibtex Entry
@article = { menczer99b,
author = "Filippo Menczer and Richard K. Belew",
title = "Adaptive Retrieval Agents: Internalizing Local Context and Scaling up to the Web",
journal = "Machine Learning",
volume = "39",
number = "2/3",
publisher = "Kluwer Academic Publishers, Boston",
pages = "203--242",
year = "2000",
url = "citeseer.nj.nec.com/menczer99adaptive.html"
}