Learning to Win

Repositories of standard-format datasets and analysis tasks defined on them have critically contributed to machine learning investigations, but their access has led to some ideological stagnation. While they allow researchers to conduct comparative empirical analyses of supervised learning algorithms on a massive scale, this led to many dull contributions on small improvements to knowledge-poor algorithms, which became easier to publish than more ambitious investigations. For example, evaluating knowledge-intensive learning approaches on challenging tasks from virtual simulation environments is far more difficult.

We have developed TIELT (Testbed for Integrating and Evaluating Learning Techniques) to address this problem. By integrating their (e.g., machine learning) system with TIELT, researchers will gain access to previously integrated simulators and comparison systems, which they can test versus their own on tasks they select using the evaluation methodology they encode. TIELT, a freely available and supported tool (http://nrlsat.ittid.com), is maturing through the efforts of industry and academic partners. I will describe our initial stress test of TIELT (Aha et al., 2005), which involved a case-based approach for learning to select player actions to win a real-time strategy game. I'll also describe our plans for using it to support DARPA challenge problems, explain its role in future gaming competitions, and encourage its use (e.g., in class projects). This work is being performed in conjunction with Matthew Molineaux (ITT Industries) and Marc Ponsen (University of Maastricht, The Netherlands).

Reference: Aha, D.W., Molineaux, M., & Ponsen, M. (2005). Learning to win: Case-based plan selection in a real-time strategy game. Proceedings of the Sixth International Conference on Case-Based Reasoning (pp. 5-20). Chicago, IL: Springer. (Winner, Best Paper Award)