The way a multi-agent modeller represents an agent not only affects its ability to reason about agents but also the interpretability of its representation space as well as its efficacy on future downstream tasks. We utilize and repurpose metrics from the field of representation learning to specifically analyze and compare multi-agent modellers that build real-valued vector representations of the agents they model. By generating two datasets and analyzing the representations of multiple LSTM- or transformer-based modellers with various embedding sizes, we demonstrate that representation metrics provide a more complete and nuanced picture of a modeller's representation space than an analysis based only on performance. We also provide insights regarding LSTM- and transformer-based representations. Our proposed metrics are general enough to work on a wide variety of modellers and datasets.

This dissertation demonstrates, in a non-semantic information-theoretic framework, how the principles of 'maximisation of relevant information' and 'information parsimony' can guide the adaptation of an agent towards agent-agent interaction. Central to this thesis is the concept of digested information; I argue that an agent is intrinsically motivated to a.) process the relevant information in its environment and b.) display this information in its own actions. From the perspective of similar agents, who require similar information, this differentiates other agents from the rest of the environment, by virtue of the information they provide. This provides an informational incentive to observe other agents and integrate their information into one's own decision making process. This process is formalized in the framework of information theory, which allows for a quantitative treatment of the resulting effects, specifically how the digested information of an agent is influenced by several factors, such as the agent's performance and the integrated information of other agents. Two specific phenomena based on information maximisation arise in this thesis. One is flocking behaviour similar to boids that results when agents are searching for a location in a girdworld and integrated the information in other agent's actions via Bayes' Theorem. The other is an effect where integrating information from too many agents becomes detrimental to an agent's performance, for which several explanations are provided.