The Doubly-Bounded Rationality of an Artificial Agent and its Ability to Represent the Bounded Rationality of a Human Decision-Maker in Policy-Relevant Situations

ABSTRACT

This article introduces two tools aimed at improving our understanding of the relationship between human and artificial rationality and helping us identify agents that are false positives or negatives. The first is a framework that systematically exposes where and how discrepancies between human and artificial rationalities can arise. The second is a test that utilises the insight gained from applying the framework in testing the ability of an artificial agent to represent human decision-making. To demonstrate the usefulness of the test, the article describes its application in testing the ability of a set of Individual Evolutionary Learning agents to represent human decision-making in a social psychology experiment, called the Voluntary Contributions Mechanism. In contrast to the results of a prior test that relied on a behaviour-based method, the results of this test show that the ability of these artificial agents to replicate the behaviour of their human counterparts is not a reliable indicator of their ability to represent their decision-making. The article then uses insight from the test to suggest how to improve the ability of Individual Evolutionary Learning agents to represent human decision-making in the Voluntary Contributions Mechanism.

KEYWORDS:

• Behaviour
• cognition
• knowledge
• bounded rationality
• doubly-bounded rationality
• testing

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

1. These reasons are not unique to just modelling BR. For example, Higgins et al. (2003) identified three types of uncertainties – model, parameter, and inherent – that are generally present in the process of modelling ecological forecasts (e.g. plant migration rates) and that are along similar lines as the three reasons mentioned above. They find these uncertainties stemming from a modeller: (a) having an incomplete understanding of the ecological system being modelled, (b) working with an incomplete data sample for the system being modelled, and/or (c) facing limitations in statistical tools. Similarly, MacLeod and Nersessian (2018) identified deficiency in data and understanding as a major challenge in modelling integrated biological systems.

2. The unboundedness of instrumental rationality is a fundamental assumption in game (and more broadly choice) theory and is represented in the respective disciplines using von Neumann and Morgenstern’s (1944) four axioms (completeness, transitivity, continuity, and independence). While the unbounded epistemic rationality remains a fundamental assumption in classic game theory, it has been relaxed in the more recently developed behavioural game theory (Camerer, 2003), which allows for boundedness of agent rationality, as long as it stems from epistemic irrationality alone.

3. Such internal constraints may not only be situation-specific, inherent generally to what Lake et al. (2017) referred to as the core set of ingredients for human intelligence, and may also be individual-specific (Stanovich & West, 1998), inherent to an individual’s unique level of intelligence.

4. Identifying biases in behaviour is currently an active area of research (Wilke & Mata, 2012). While peer-reviewed literature on biases provides an abundance of lists with examples of biases (e.g. Stanovich, 2016; Tversky & Kahneman, 1974; Wilke & Mata, 2012), Wikipedia’s non-peer-reviewed list of nearly 200 mostly peer-reviewed biases currently both includes and dwarfs most (if not all) of the other lists.

5. This is because the jury is still out on whether mental models and heuristics are sufficiently appropriate constructs for organising the knowledge a BR decision maker uses in their decision-making.

6. Interestingly, such cases are possible even when von Neumann and Morgenstern’s (1944) four axioms hold, because the axioms address only the instrumental and not the epistemic dimension of rationality.

7. Other behaviour patterns commonly observed in VCM experiments (e.g. Chaudhuri, 2011; Holt & Laury, 2008; Ledyard, 1995) are not discussed in Arifovic and Ledyard (2012) and are therefore unavailable for comparison.