1 Acting Deceptively: Providing Robots with the Capacity for Deception Alan R. Wagner Georgia Institute of Technology and Georgia Tech Research Institute 85 5th Street, Room S27, NW, Atlanta, GA. 30308 1-404-894-9311 [email protected], [email protected] Ronald C. Arkin Georgia Institute of Technology 85 5th Street, S27, NW, Atlanta, GA. 30308 1-404-894-9311 [email protected] Abstract Deception is utilized by a variety of intelligent systems ranging from insects to human beings. It has been argued that the use of deception is an indicator of theory of mind [2] and of social intelligence [4]. We use interdependence theory and game theory to explore the phenomena of deception from the perspective of robotics, and to develop an algorithm which allows an artificially intelligent system to determine if deception is warranted in a social situation. Using techniques introduced in [1], we present an algorithm that bases a robot’s deceptive action selection on its model of the individual it’s attempting to deceive. Simulation and robot experiments using these algorithms which investigate the nature of deception itself are discussed. Keywords: deception, game theory, interdependence theory, interaction, hide-and-seek, theory of mind. 1. Introduction Deception has a long and deep history with respect to the study of intelligent systems. Biologists and psychologists argue that deception is ubiquitous within the animal kingdom and represents an evolutionary advantage for the deceiver [5]. Primatologists note that the use of deception serves as an important potential indicator of theory of mind [2] and social intelligence [4]. Researchers in these fields point to numerous examples of deception by non-human primates. From a roboticist’s perspective, the use of deception and the development of strategies for resisting being deceived are important topics of study especially with respect to the military domain [7].2 But what is deception? McCleskey notes that deception is a deliberate action or series of actions brought about for a specific purpose [8]. Whaley recognizes that deception often includes information provided with the intent of manipulating some other individual [9]. Ettinger and Jehiel offer a related definition describing deception as, “the process by which actions are chosen to manipulate beliefs so as to take advantage of the erroneous inferences [10].” This definition has clear ties to game theory but does not relate to many of the passive, unintentional examples of deception found in biology. We adopt a definition for deception offered by Bond and Robinson that encompasses conscious and unconscious, intentional and unintentional acts of deception. These authors describe deception simply as a false communication that tends to benefit the communicator [5]. This paper investigates the use of deception by autonomous robots. We focus on the actions, beliefs and communication of the deceiver, not the deceived. Specifically, our central thesis is that modeling of the individual to be deceived is a critical factor in determining the extent to which a deceptive behavior will be effective. In other words, a robot must have specific knowledge about the individual that it is attempting to deceive—the mark—in order for the deceptive action to be effective. It is worth noting that a deceiver’s knowledge of the mark need not be explicit. The exploration of this thesis is important both for the creation of deceptive robots and for developing a better understanding of the nature of deception itself. Consider, for example, the use of camouflage. Camouflage is the use of natural or artificial material to allow an otherwise visible object to remain indiscernible from the surrounding environment [7]. The act of camouflaging assumes that the mark has specific3 perceptual characteristics, such as color vision. A robot that relies on infrared, for example, will not be deceived by color-based camouflaging. We will use the following example to illustrate our ideas: a valuable robotic asset operates at a military base. The base comes under attack and is in danger of being overrun. If the robot is discovered by the attackers then they will gain valuable information and hardware. The robot must hide and select a deceptive strategy that will reduce the chance that it will be encountered. Throughout this article we will use this running example to explain portions of the theoretical underpinnings of our approach as well as to develop experiments based on the example. The remainder of this paper begins by first summarizing relevant research. Next, we use game theory and interdependence theory to reason about the theoretical underpinnings of deception and to develop preliminary algorithms for the effective use of deception on a robot. Finally, we present a series of experiments which attempt to investigate the veracity of our thesis. The article concludes with a discussion of these results including directions for future research and the ethical implications of our research. 2. Related Work Game theory has been extensively used to explore the phenomena of deception. As a branch of applied mathematics, game theory focuses on the formal consideration of strategic interactions, such as the existence of equilibriums and economic applications [11]. Signaling games, for example, explore deception by allowing each individual to send signals relating to their underlying type [12]. Costly versus cost-free signaling has been used to determine the conditions that foster honesty. Floreano et al. found that4 deceptive communication signals can evolve when conditions conducive to these signals are present [13]. These researchers used both simulation experiments and real robots to explore the conditions necessary for the evolution of communication signals. They found that cooperative communication readily evolves when robot colonies consist of genetically similar individuals. Yet when the robot colonies were genetically dissimilar and evolutionary selection of individuals rather than colonies was performed, the robots evolved deceptive communication signals, which, for example, compelled them to signal