Back
to indexFilippo Menczer and Richard K. Belew
Summary
This paper presents the ARACHNID architecture, with emphasis on the textual environment the agent operates in. The word topology of the environment is defined to be the similarity between topics that would appear close together in the search space. The topology provides the agent with context, from which it can infer differences between 'rock music' and 'rock climbing'. ARACHNID incorporates the topology into its search by choosing which link it should follow next, and reinforcement learning upon success or failure of that link. Agents mutate or clone themselves within the classical framework of evolutionary computation. Search length is defined to be the distance to the first hit page from the starting point.
Experimental evidence and frequent citations lend credibility to this work, as a foundation upon which future agents may be built. I would possibly extend the work by incorporating repeated use of neural net weights of the best agents in a series, with the hope to generate an even higher quality agent over time. The adaptability of these agents with respect to network resources is an area that needs to be explored further as well.
Methods
Each agent maintains a keyword list, which is weighted and modified throughout the agent's lifetime. Occurrences of words in the keyword list is used as input to a neural network that aids the agent in determination of which link to follow next, through Q-learning. Energy thresholds determine if an agent may clone or mutate itself, and if it should be killed. Relevance is determined through term frequency x inverse document frequency measurements.
Keywords
reinforcement learning, ARACHNID, word topology, genotype, Q-learning, recombination
Rating
8
Bibtex Entry
@inproceedings = { menczer98b,
author = "Filippo Menczer and Richard K. Belew",
title = "Adaptive Information Agents in Distributed Textual Environments",
booktitle = "Proceedings of the 2nd International Conference on Autonomous Agents (Agents'98)",
month = "~9--13,",
publisher = "ACM Press",
address = "New York",
editor = "Katia P. Sycara and Michael Wooldridge",
isbn = "0-89791-983-1",
pages = "157--164",
year = "1998",
url = "citeseer.nj.nec.com/menczer98adaptive.html"
}