Group living benefits foraging individuals by improving their survival through passive risk dilution by sheer numbers and increasingly active processes, ranging from cue transmission to alarm calling. Focusing on the lower end of this range: an involuntary visual cue can be given by a fleeing action, leading to cue transmission within a group. Our model is a bottom-up model of foragers as agents embodied in a simple environment, with only assumptions about basic living competences, valid for a wide range of species, leading to conclusions about benefits of group living without calling on higher cognition.We use an agent-based, spatially explicit C++-model to investigate the effect of predator disturbances, to which always fleeing is the appropriate reaction. To make the environment more realistic and to include not only predators, we added harmless passers-by that cause false, unnecessary, fleeing. We set out to investigate whether adaptive behaviour could improve outcomes: We investigated whether two common subconscious behaviours can mitigate the detrimental effects of false fleeing. The first is “experience gain”, a sensory change; the second is “fear updating,” an emotional change.To implement these behaviours we needed to: 1) model the anti-predator behaviour chain of “detection, recognition, and response” for which we added a recognition phase, 2) handle the combined probabilities of the detection and recognition sigmoids, and 3) define the initial settings for the three fear levels we added: one each for predator and passer-by and one for the environment, needed for the response in case of detection without recognition. These modelling decisions are shown to be very important. Situations with new predators or newly released groups can be mapped to these initial settings, distinguishing whether the forager group encounters a familiar predator type or a novel one, for instance, giving insight into what is most important for forager groups.

In formal epistemology, a variety of probability-based coherence measures have been proposed that provide a quantitative formal representation of the coherence of a set of information pieces. While research has long focused on whether coherence measures are truth-conducive, the truth-conduciveness of coherence measures has so far been evaluated in static settings only: Coherence provides assessments about the truth of incoming information, but does not actively guide decisions to believe or discard pieces of information. In this paper, we propose to assess the truth-conduciveness of coherence measures with respect to their ability to lead agents to select true information and form correct beliefs in a dynamic iterative setting. At every time step, an agent receives a number of noisy signals about the actual truth values of a finite set of atomic propositional variables. The agent uses a coherence measure to decide which signals to trust and which to discard. By repeatedly picking signals that maximise the coherence of the propositions they currently believe to be true, the agent tries to select truthful signals and learn the correct truth-value assignment for the atomic variables. The contribution of this paper is three-fold. First, we propose a computational model to assess the truth-tracking abilities of different coherence measures. Second, using computational simulations, we compare a number of widely discussed coherence measures from the novel standpoint of our iterated data-collection setting: We show that, when signals are not too noisy, agents who employ the Glass-Olsson relative overlap measure outperform agents employing all other tested measures, and that all measures become progressively worse at leading agents towards the truth as signals degrade. Finally, we discuss how coherence affects the emergence of different dynamics and attitudes in belief revision.

Theory of mind (ToM) is the ability to attribute and reason about unobservable mental states of others, such as knowledge, beliefs, and intentions. ToM is needed in social interactions, negotiations, cooperation, and deception, but there is a limit to the recursive depth of ToM attributions humans can make. Researchers in epistemic logic classically assume that agents have common knowledge of honesty and perfect rationality, whereas humans lie and have ToM limitations. In our present work, we bridge this gap by developing a sound and strongly complete dynamic epistemic logic that models both lying and human ToM limitations. The design of our logic is inspired by findings from prior behavioural research. Our logic makes use of action models to model belief change caused by lying and other types of announcements. Using theoretical examples, we show how a listener deals with lies that she cannot understand due to her ToM limits.

Multiagent systems bring together agents that represent different users with possibly different concerns. When interacting to make decisions, conflicts occur. A well-known case is with privacy. Agents often need to manage the privacy of content that belong to multiple users, such as sharing group pictures on social media. When agents have different expectations on how the content should be shared, multi-party privacy conflicts can arise. How should we design agents to deal with such conflicts? We have studied an empirical user study to understand the effect of group dynamics in various multi-party privacy settings. Our findings show that as users' beliefs and knowledge about others evolve, privacy expectations shift as well. Inspired by this, we propose computational agents that mimic a human-inspired Theory of Mind (ToM) model to help their users preserve their privacy in multi-party privacy conflicts. The agents can express empathy when others are in need but can also fight for their own privacy. We evaluate our approach in multiagent simulations with varying decision-making strategies. Our results demonstrate that ToM-enabled agents improve privacy preservation for all parties, and even more when their understanding of others is dynamically updated through learning.

The principle of minimal change in belief revision theory requires that, when accepting new information, one keeps one's belief state as close to the initial belief state as possible. This is precisely what the method known as minimal revision does. However, unlike less conservative belief revision methods, minimal revision falls short in learning power: It cannot learn everything that can be learned by other learning methods. We begin by showing that, despite this limitation, minimal revision is still a successful learning method in a wide range of situations. Firstly, it can learn any problem that is finitely identifiable. Secondly, it can learn with positive and negative data, as long as one considers finitely many possibilities. We then characterize the prior plausibility assignments (over finitely many possibilities) that enable one to learn via minimal revision, and do the same for conditioning and lexicographic upgrade. Finally, we show that not all of our results still hold when learning from possibly erroneous information.

Philosopher: Our project is drawing to an end, and our previous discussion has made clear that it will not be so easy for the social software enterprise to address, let alone cure, the ills of society. To get the discussion going, I would like to start today with a quote that I found in a collection of talks from the physicist Richard Feynman. It is from one of the pieces in The Pleasure of Finding Things Out, a digest of a talk on “The Value of Science”. The talk starts like this:

The search for effective collaboration between humans and computer systems is one of the biggest challenges in Artificial Intelligence. One of the more effective mechanisms that humans use to coordinate with one another is theory of mind (ToM). ToM can be described as the ability to `take someone else's perspective and make estimations of their beliefs, desires and intentions, in order to make sense of their behaviour and attitudes towards the world'. If leveraged properly, this skill can be very useful in Human-AI collaboration. This introduces the question how we implement ToM when building an AI system. Humans and AI Systems work quite differently, and ToM is a multifaceted concept, each facet rooted in different research traditions across the cognitive and developmental sciences. We observe that researchers from artificial intelligence and the computing sciences, ourselves included, often have difficulties finding their way in the ToM literature. In this paper, we identify four common misconceptions around ToM that we believe should be taken into account when developing an AI system. We have hyperbolised these misconceptions for the sake of the argument, but add nuance in their discussion. The misconceptions we discuss are: (1) "Humans Use a ToM Module, So AI Systems Should As Well". (2) "Every Social Interaction Requires (Advanced) ToM". (3) "All ToM is the Same". (4) "Current Systems Already Have ToM". After discussing the misconception, we end each section by providing tentative guidelines on how the misconception can be overcome.

Tunnel vision and confirmation bias can lead to miscarriages of justice. A way to avoid tunnel vision is to consider your evidence in light of more than one scenario. Alternative scenarios allow us to consider how probable each scenario is, compared to the other considered scenarios. Bayesian Networks have been proposed as a formal method for reasoning about the probability of scenarios. Specifically, alternative scenarios were modelled using Bayesian networks with a constraint node, which ensures mutual exclusivity. However, the performance of these methods in situations where not all possible alternative scenarios are modeled, the non-exhaustive setting, has not been investigated. Since it is impossible to explicitly cover everything that could possibly have happened in a model, it is important to know how these methods handle non-exhaustiveness. We evaluate four methods using an agent-based model that simulates an environment in which a crime could occur. Taking this as the ground truth, we compare different Bayesian network modeling methods on five aspects related to the quality of the representation of the ground truth as well as computational performance. We find that some methods result in disparities between the ground truth and the predicted posterior probabilities for the scenarios in a non-exhaustive setting. In an exhaustive setting, the proposed methods perform well. The construction approach that models scenarios in terms of conjunctions of events performs well in both settings.

Theory of mind refers to the human ability to reason about the mental content of other people, such as their beliefs, desires, and goals. People use their theory of mind to understand, reason about, and explain the behaviour of others. Having a theory of mind is especially useful when people collaborate, since individuals can then reason on what the other individual knows as well as what reasoning they might do. Similarly, hybrid intelligence systems, where AI agents collaborate with humans, necessitate that the agents reason about the humans using computational theory of mind. However, to try to keep track of all individual mental attitudes of all other individuals becomes (computationally) very difficult. Accordingly, this paper provides a mechanism for computational theory of mind based on abstractions of single beliefs into higher-level concepts. These abstractions can be triggered by social norms and roles. Their use in decision making serves as a heuristic to choose among interactions, thus facilitating collaboration. We provide a formalization based on epistemic logic to explain how various inferences enable such a computational theory of mind. Using examples from the medical domain, we demonstrate how having such a theory of mind enables an agent to interact with humans effectively and can increase the quality of the decisions humans make.

With the growing integration of chatbots, automated writing tools, game AI and similar applications into human society, there is a clear demand for artificially intelligent systems that can successfully collaborate with human partners. This requires overcoming not only physical and communicative barriers, but also those of fundamental understanding: Machines do not see and understand the world in the same way as humans do. We introduce the concept of ‘Common Ground’ (CG) as a possible solution. Using a model inspired on a collaborative card game known as ‘The Game’, we study agents that are instantiated to use different strategies, i.e., they each ‘see’ the model world in a different way. Agents work towards a joint goal that is easy to understand but complex to attain, requiring them to constantly anticipate their partner, which is classically seen as a task requiring active perspective modelling using a form of Theory of Mind. We show that agents achieving Common Ground increase their joint performance, while the need to actively model each other decreases. We discuss the implications of this finding for interaction between computational agents and humans, and suggest future extensions of our model to study the benefits of CG in hybrid human-agent settings.

Empowering artificially intelligent agents with capabilities that humans use regularly is crucial to enable effective human-agent collaboration. One of these crucial capabilities is the modeling of Theory of Mind (ToM) reasoning: the human ability to reason about the mental content of others such as their beliefs, desires, and goals. However, it is generally impractical to track all individual mental attitudes of all other individuals and for many practical situations not even necessary. Hence, what is important is to capture enough information to create an approximate model that is effective and flexible. Accordingly, this paper proposes a computational ToM mechanism based on abstracting beliefs and knowledge into higher-level human concepts, called abstractions, similar to the ones that guide humans to effectively interact with each other (e.g., trust). We develop an agent architecture based on epistemic logic to formalize the computational dynamics of ToM reasoning. We identify important challenges regarding effective maintenance of abstractions and accurate use of ToM reasoning and demonstrate how our approach addresses these challenges over multiagent simulations.

This paper introduces a comprehensive logical framework to reason about threshold-driven diffusion and threshold-driven link change in social networks. It considers both monotonic dynamics, where agents can only adopt new features and create new connections, and non-monotonic dynamics, where agents may also abandon features or cut ties. Three types of operators are combined: one capturing diffusion only, one capturing link change only, and one capturing both at the same time. We first characterise the models on which diffusion of a unique feature and link change stabilise, whilst discussing salient properties of stable models with multiple spreading features. Second, we show that our operators (and any combination of them) are irreplaceable, in the sense that the sequences of model updates expressed by a combination of operators cannot always be expressed using any other operators. Finally, we analyse classes of models on which some operators can be replaced.