We show that adaptive agents on the Internet can learn to exploit bidding agents who use a (limited) number of fixed strategies. These learning agents can be generated by adapting a special kind of finite automata with evolutionary algorithms (EAs). Our approach is especially powerful if the adaptive agent participates in frequently occurring micro-transactions, where there is sufficient opportunity for the agent to learn online from past negotiations. More in general, results presented in this paper provide a solid basis for the further development of adaptive agents for Internet applications.

Optimization (acm G.1.6), Learning (acm I.2.6), Problem Solving, Control Methods, and Search (acm I.2.8)
Learning and adaptive systems (msc 68T05), Problem solving (heuristics, search strategies, etc.) (msc 68T20), 2-person games (msc 91A05), Noncooperative games (msc 91A10), Games in extensive form (msc 91A18), Multistage and repeated games (msc 91A20), Evolutionary games (msc 91A22), Rationality, learning (msc 91A26)
Software (theme 1), Logistics (theme 3), Energy (theme 4)
Software Engineering [SEN]
Intelligent and autonomous systems

van Bragt, D.D.B, & La Poutré, J.A. (2002). Why agents for automated negotiations should be adaptive. Software Engineering [SEN]. CWI.