Computational models of decision making

This chapter presents a connectionist or artificial neural network approach to decision making. An essential idea of this approach is that decisions are based on the accumulation of the affective evaluations produced by each action until a threshold criterion is reached. This type of sequential sampling process forms the basis for decision making in a wide variety of other cognitive tasks such as perception, categorization, and memory. We apply these concepts to several important preferential choice phenomena, including similarity effects, attraction effects, compromise effects, loss aversion effects, and preference reversals. These analyses indicate that a relatively complex model of an individual’s choice process reveals a relatively simple representation of the individual’s underlying value structure. Computational Models 3 What are computational models of cognition? In his classic book on computational vision, Marr (1982) proposed three levels of theories about cognitive systems. At the highest level, theories aim to understand the abstract goals a sys tem is trying to achieve; at an intermediate level, theories are designed to explain the dynamic processes used to achieve the top level goals; and at the bottom level, theories attempt to describe the neurophysiologic substrate of the second level. Judgment and decision-making researchers have generally been concerned with theorizing at the higher and more abstract levels. From this higher point of view, explanations based on principles such as context dependent weights, loss aversion, and anchoring-adjustment are considered satisfactory. This chapter presents arguments for viewing decision making from the perspective of a lower level microanalysis. By doing so, we can try to answer deeper questions such as: why decision weights change across contexts, why people are loss averse, and why anchors are more influential than adjustments. Computational models are constructed from simple units that conform to a small number of elementary principles of cognition, but a large number of these simple units are connected together to form a dynamical system. Although the properties of the individual units are simple, the emergent behavior of the ensemble becomes fairly complex. Computational models appear in a variety of forms, but this chapter focuses on a class known as artificial neural networks, connectionist networks, or parallel distributed processing systems (see Grossberg, 1988; and Rumelhart & McClelland, 1986, for general overviews of these models). This class of computational models is designed to form a bridge that mediates between the neural and behavioral sciences. Computational Models 4 How does the brain make decisions? A decade ago, the brain was an impenetrable black box, but with recent advances in neuroscience, we can start to look inside. It is informative to point out a conclusion arising from converging evidence obtained through neuroscience research on decisionmaking. Neuroscientists have examined decision-making processes in the brains of Macaque monkeys using single cell recording techniques (for reviews, see Gold & Shadlen, 2001, 2002; Platt, 2002; Schall, 2001), as well as from the brains of humans using evoked response potentials (Gratten, Coles, Sirevaag, & Donchin, 1988). A simple but important conclusion from this work is that decisions in the brain are based on the dynamic accumulation of noisy activation for each action, and the action whose activation first exceeds the threshold is chosen. This process is illustrated in Figure 1, for three actions, with each trajectory representing the cumulative activation (i.e., preference state) for an action. The horizontal axis represents deliberation time and the vertical axis indicates the activation for each action at each moment in time. In this figure, action A reaches the threshold first, and is chosen at time T = 425. Computational Models 5 Figure 1: The decision process for a choice among three actions 0 100 200 300 400 500 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Deliberation Time P re fe re nc e S ta te Threshold Bound