Learning cooperative behavior for the shout-ahead architecture

Article type: Research Article

Authors: Paskaradevan, Sanjeev | Denzinger, Jörg^; | Wehr, Daniel

Affiliations: Department of Computer Science, University of Calgary, 2500 University Drive NW, Calgary, AB, Canada, T2N 1N4. E-mail: [email protected], {denzinge,dkwehr}@ucalgary.ca

Note: [] Corresponding author. E-mail: [email protected].

Abstract: We present an agent architecture and a hybrid behavior learning method for it that allows the use of communicated intentions of other agents to create agents that are able to cooperate with various configurations of other agents in fulfilling a task. Our shout-ahead architecture is based on two rule sets, one making decisions without communicated intentions and one with these intentions, and reinforcement learning is used to determine in a particular situation which set is responsible for the final decision. Evolutionary learning is used to learn these rules. Our application of this approach to learning behaviors for units in a computer game shows that the use of shout-ahead using only communicated intentions in the second rule set substantially improves the quality of the learned behavior compared to agents not using shout-ahead. Also, allowing for additional conditions in the second rule set can either improve the quality or worsen it, based on what type of conditions are used.

Keywords: Cooperative systems, reinforcement learning, evolutionary learning, communication, shout ahead

DOI: 10.3233/WIA-140298

Journal: Web Intelligence and Agent Systems: An International Journal, vol. 12, no. 3, pp. 309-324, 2014