What Do You Want From Me? Adapting Systems to the Uncertainty of Human Preferences

Abstract

Autonomous systems, like drones and self-driving cars, are becoming part of our daily lives. Multiple people interact with them, each with their own expectations regarding system behaviour. To adapt system behaviour to human preferences, we propose and explore a game-theoretic approach. In our architecture, autonomous systems use sensor data to build game-theoretic models of their interaction with humans. In these models, we represent human preferences with types and a probability distribution over them. Game-theoretic analysis then outputs a strategy, that determines how the system should act to maximise utility, given its beliefs over human types. We showcase our approach in a search-and-rescue (SAR) scenario, with a robot in charge of locating victims. According to social psychology, depending on their identity some people are keen to help others, while some prioritise their personal safety. These social identities define what a person favours, so we can map them directly to game-theoretic types. We show that our approach enables a SAR robot to take advantage of human collaboration, outperforming non-adaptive configurations in average number of successful evacuations. CCS CONCEPTS• Computer systems organization → Robotics; • Human-centered computing → Collaborative interaction.

“Are we in this together?”: embedding social identity detection in drones improves emergency coordination

“…Based on this information, the identity-informed drone can decide whether a victim is likely to be helped by fellow survivors or whether assistance from emergency responders is needed, and, thus, what action needs to be taken. In the following, we provide an overview of the theoretical and practical drivers of our investigation and describe the architecture developed in previous research ( Gavidia-Calderon et al, 2022 ), which allows for identity adaptation. We, then, describe our proposed methodology for this adaptation based on survivors’ linguistic features.…”

“…To be able to accurately detect and make decisions in line with a survivor’s identity, the autonomous system needs a self-adaptive architecture that includes an identity classifier-a function to computationally detect an identity and quantify it to inform decision making. Such an architecture was designed in previous work ( Gavidia-Calderon et al, 2022 ) using a MAPE-K framework for self-adaptive systems ( Kephart and Chess, 2004 ). Under this framework, the autonomous system’s components were organized within four elements: Monitor, Analyze, Plan, and Execute that operate over a shared Knowledge .…”

“…If the drone detects a first responder close to the victim, adaptation for providing help is not required. If the drone detects another survivor close to the victim, adaptation is required ( Gavidia-Calderon et al, 2022 ).…”

“…If a personal identity is detected, the drone needs to request help from a first responder remotely (hence, the adaptation). In Gavidia-Calderon et al (2022) , these interactions between a victim who requires help, a survivor in close proximity, and the drone are represented using Game Theory. Game theoretic models are models that represent interactions of conflict or cooperation between rational decision makers who adopt certain strategies to perceive an impact, i.e., maximize their payoff ( Tadelis, 2013 ).…”

“…Then, the Plan element calculates the Bayesian Nash equilibria, namely a predicted strategy of the survivor and the response that the drone should have to this strategy. The Execute element translates these strategies into robot commands, i.e., asking for help from the survivor or the first responder (see Gavidia-Calderon et al, 2022 for a more detailed explanation of the identity-adaptive architecture that was adopted here). The focus of this paper is on our methodological approach for the identity classifier of this architecture based on linguistic data, which is presented next.…”