TL;DR#
The paper investigates the concept of agency in artificial intelligence and reinforcement learning. Current definitions of agency treat it as a binary property—a system either possesses agency or it does not. This approach presents challenges because determining agency often requires subjective judgments. For example, it is difficult to determine whether a thermostat, a rock, or a robot is an agent because the answer depends on how one defines the terms of ‘agency’ and what one counts as relevant.
The paper proposes a novel framework that addresses this issue. The authors argue that agency is fundamentally frame-dependent. This means that determining whether a system is an agent is dependent on the chosen frame of reference, which is formed by commitments about aspects of the system such as its boundary, goals and so on. The authors use this frame-dependent view to support their argument that the four components of agency—individuality, source of action, goal-directedness and adaptivity—are all frame-dependent. The paper’s main contribution is to shift the focus from binary, absolute concepts of agency to a more nuanced, relative understanding. This has implications for designing and interpreting reinforcement learning algorithms and the study of agency in general.
Key Takeaways#
Why does it matter?#
This paper is important because it challenges the fundamental understanding of agency in artificial intelligence and reinforcement learning. By introducing the concept of frame-dependence, it opens new avenues for research in defining and measuring agency, potentially leading to more robust and adaptable AI systems. It also connects with current trends in causal reasoning and system identification within AI.
Visual Insights#
🔼 This figure illustrates Barandiaran et al.’s (2009) four-part account of agency, which posits that a system demonstrates agency if it satisfies four conditions: individuality (it has a boundary distinguishing it from the environment), it is the source of its own actions, it is goal-directed (it has goals or norms), and it is adaptive (it adjusts its actions based on experiences). This framework is crucial to the paper’s argument about frame-dependence, as each of these four properties is shown to be relative to the chosen frame of reference.
read the caption
(a) A Four-Part Account of Agency
Full paper#
