Understanding, Replicating, and Leveraging Dynamics of Bidder Behavior in Continuous Combinatorial Auctions

Studying bidder behavior in combinatorial auctions is an integral aspect for the design of effective and feasible mechanisms, since users' preferences and behaviors are intricately linked to the design of information systems. In our previous work, we studied and analyzed bidder behavior in laboratory based combinatorial auctions and observed its effect on auction performance metrics (e.g., auctioneer revenue and efficiency).  Experiments, though, are usually costly and the number of treatment variables that can be analyzed are limited. More importantly, it is not possible to control a priori for the different types of bidder behavior. Hence, only few possible compositions of competitive behaviors can be observed, and the effect of more diverse types of competition on auction dynamics and outcomes could not be necessarily observed or predicted.  Similar to current scientific paradigms, agent-based computer simulation is becoming a social-science research method that complements analytical, empirical, and experimental approaches to investigate and identify potential interesting settings that could then be explicitly tested and explored in laboratory or real-world setting. Simulations also make it possible to investigate scenarios and study phenomena that are difficult to analyze analytically and/or are difficult (sometimes impossible) to test in experimental settings. The design of simulation models can be based on, or informed by, theoretically derived properties and/or empirical data from experiments or real world cases. These approaches can be used in combination with each other, for example, analytical models can inform the better design of experiments and computational simulations can be based on theoretical models.We use an agent-based modelling approach that enables us to create, analyze, and experiment with simulation models composed of bidding agents that interact within a combinatorial auction environment. Agent-based models are a class of computational simulation, which are similar to mathematical modeling in terms of rigor but better suited for situations when agents are autonomous and heterogeneous, when there are complex interactions between agents, and when lower-level actions and interactions can lead to the emergence of system-level structure. We use a data driven approach to develop novel software agents that replicate human bidder behavior, based on experimental data from continuous combinatorial auctions (henceforth called CCAs). These bidding agents are used in auction simulations to explore all possible auction scenarios and different bidder type compositions (i.e., different competition types) that may not be easily encountered when running combinatorial auctions in experimental lab environments. We can leverage these agents to control for the type of competition and analyze the dynamics of bidder behavior and emerging auction outcomes such as revenue and efficiency. This is the first study that uses an agent-based modeling approach to simulate human bidder behavior in CCAs. Findings from our computational agent based simulations allow for bottom-up theorizing and deeper understanding how individual bidders’ behavior interact and lead to emergent auction outcomes. It also contributes to the better design of smart markets by providing a better understanding of bidder behavior in CCAs, and uncovering the effect of competition on bidder behavior and auction outcomes.