ALF

Advanced Learning Framework

Competence Center: Information Retrieval and Machine Learning  
Contact: Prof. Dr. Sahin Albayrak, Dipl.-Inf. Christian Scheel
Partners: Prof. Obermayer, NI, TU Berlin  

 

ALF employs methods for distributed, agent-based information filtering. The cornerstone property of ALF lies in the usage of different filtering communities for different databases. Each filtering community has different filtering agents and one manager agent that is responsible for coordination and cooperation tasks. While cooperation mechanisms are contained in the delegation of queries between communities, coordination mechanisms take place inside each filtering community. More importantly, coordination mechanisms ensure

  • that filtering agents are not activated, when the necessary resources are not available,
  • that filter managers learn, based on the received feedback, which filtering agent is suitable for what kinds of tasks, and
  • that in case of failure an alternative filtering agent is selected which finally increases the robustness. 

Goals

  • A software infrastructure for the realization of manager and filtering agents will be developed on the basis of the JIAC IV framework.
  • Manager strategies for efficient creation of communities, delegation of queries and cooperation with other communities will be realized.
  • A palette of filtering techniques will be implemented within different filtering agents. These strategies should be able to evaluate the quality of results both before and after processing. Additionally, they should be capable to estimate the amount of resources needed for processing the query.
  • The system behavior and the different manager strategies will be investigated. An empirical evaluation with both synthetic data and scientific papers will be performed after the implementation.

Implementation

The experience that was gained in the URLAUB and PIA projects was the motivation for enhancing their architectures in order to combine the flexibility of multi-agent systems and the adaptability of machine learning techniques. Therefore, different machine learning techniques will be encapsulated in filtering agents that become capable of adapting to user preferences. A manager agent builds a so-called community for a particular information source, decides which filtering agent will process a particular query, and cooperates with other communities. Filtering agents are chosen based on the expected quality of their filter results. Because the manager agent can take care of the availability of resources, an excellent scalability is ensured. Finally, it should be determined under which circumstances such a heterogeneous system with learning abilities performs better than other typical information retrieval systems.