Spatial competition of learning agents in agricultural procurement markets

Abstract

Spatially dispersed farmers supply raw milk as the primary input to a small number of large dairy-processing firms. The spatial competition of processing firms has short- to long-term repercussions on farm and processor structure, as it determines the regional demand for raw milk and the resulting raw milk price. A number of recent analytical and empirical contributions in the literature analyse the spatial price competition of processing firms in milk markets. Agent-based models (ABMs) serve by now as computational laboratories in many social science and interdisciplinary fields and are recently also introduced as bottom-up approaches to help understand market outcomes emerging from autonomously deciding and interacting agents. Despite ABMs' strengths, the inclusion of interactive learning by intelligent agents is not sufficiently matured. Although the literature of multi-agent systems (MASs) and multi-agent economic simulation are related fields of research they have progressed along separate paths. This thesis takes us through some basic steps involved in developing a theoretical basis for designing multi-agent learning in spatial economic ABMs. Each of the three main chapters of the thesis investigates a core issue for designing interactive learning systems with the overarching aim of better understanding the emergence of pricing behaviour in real, spatial agricultural markets.<br /> An important problem in the competitive spatial economics literature is the lack of a rigorous theoretical explanation for observed collusive behavior in oligopsonistic markets. The first main chapter theoretically derives how the incorporation of foresight in agents' pricing policy in spatial markets might move the system towards cooperative Nash equilibria. It is shown that a basic level of foresight invites competing firms to cease limitless price wars. Introducing the concept of an outside option into the agents' decisions within a dynamic pricing game reveals viihow decreasing returns for increasing strategic thinking correlates with the relevance of transportation costs.<br /> In the second main chapter, we introduce a new learning algorithm for rational agents using H-PHC (hierarchical policy hill climbing) in spatial markets. While MASs algorithms are typically just applicable to small problems, we show experimentally how a community of multiple rational agents is able to overcome the coordination problem in a variety of spatial (and non-spatial) market games of rich decision spaces with modest computational effort. The theoretical explanation of emerging price equilibria in spatial markets is much disputed in the literature. The majority of papers attribute the pricing behavior of processing firms (mill price and freight absorption) merely to the spatial structure of markets. Based on a computational approach with interactive learning agents in two-dimensional space, the third main chapter suggests that associating the extent of freight absorption just with the factor space can be ambiguous. In addition, the pricing behavior of agricultural processors – namely the ability to coordinate and achieve mutually beneficial outcomes - also depends on their ability to learn from each other.