Next: Comparison with previous Up: Learning Agents for Previous: Learning Agents for

The proposed approach

Drawing inspiration from the two above-mentioned fields of research, this thesis proposes the use of Interface Agents for Personalized Information Filtering. IF is an ideal and challenging environment for the use of interface agents. The information space is complex and getting more so. It is also very dynamic and the new information coming in is highly unpredictable. To add to the complexity, the user needs could themselves be changing. Adaptive agents can help users cope with this flux of information. Furthermore, IF meets the two requirements for applicability of agents - repetitive computer interaction and differences in individual preferences.

The proposed idea is to build a set of adaptive autonomous interface agents that inhabit the user's computer and are assigned the goal of being responsive to the information needs of the user. The agents can sense user feedback as well as changes in the information environment. The agents are autonomous as they can take actions relating to news filtering on the user's behalf. The agents are adaptive as they learn the preferences of the user and adapt as they change over time. The learning mechanism used by the agents is relevance feedback and genetic algorithm. The profiles used by the filtering system consist of terms which are matched with the contents of the documents i.e. the agents use cognitive filtering. The algorithm used by the agent is described in the following chapter.

The learning mechanism used in the information filtering agents is motivated by research in Genetic Algorithms and Artificial Evolution [19][18][15][6][1]. IF is effectively a dynamically changing search problem. Searching a large and changing space involves a trade-off between two objectives: (i) exploiting the currently available solution and (ii) further exploring the search space for a possibly better solution. Hill Climbing is an example of a search technique that exploits the best known alternative. However, because of this very reason, it is likely to get stuck in local maxima. Random Search, on the other hand, is an extreme case of an exploration search technique: it is unsatisfactory as it does not make use of the best solution found so far. Genetic Algorithms manage the trade-off between exploration and exploitation in a near optimal way - they exploit the solution found so far, while Crossover and Mutation operations provide a way of exploring the search space for better solutions [18].

Several experiments have demonstrated that artificial evolution is helped by individual learning [19][1]. This phenomenon is also known as the ``Baldwin effect'': if the organisms evolved are allowed to learn during their lifetime, then the evolution towards a fitter species happens much faster. This is the case because every individual is able to explore a ``patch'' of the search space (find the maximum fitness in the local neighborhood of its genotype) rather than a single point (evaluate the fitness of its own genotype).

Drawing upon the above-mentioned ideas, the information filtering agent is modeled as a population of profiles. Each profile searches for documents that match itself and recommends them to the user. The user can provide feedback for the documents recommended. User feedback causes two effects. One, it changes the fitness of the profiles. If the user provides positive (negative) feedback for a document, the fitness of the profile which retrieved that document is increased (decreased). Second, the profile is modified in response to user feedback. Thus, each profile learns during its lifetime, taking advantage of the Baldwin effect. The population as a whole continually adapts to the dynamic needs of the user.



Next: Comparison with previous Up: Learning Agents for Previous: Learning Agents for


MIT Media Lab - Autonomous Agents Group - agentmaster@media.mit.edu