Search

Machine Learning is a powerful tool for both processing and generating data. It has been demonstrated to be more efficient than humans at distinguishing hundreds of different types of images. Such classification and game-based metrics are easily quantifiable, making it simple to demonstrate improvements in efficiency and quality over previous iterations. In contrast, Generative Models (GMs) that create synthetic samples by simulating a distribution are much harder to evaluate cleanly. When looking at images, an approach known as Maximum Mean Discrepancy (MMD) has recently become quite popular, as it can non-parametrically compare samples and assign a similarity score between their relative distributions. A major flaw that MMD has is that there is no accepted approach for defining a score that would deem the generated images as similar enough to real images for them to pass an independent analysis. Introducing humans has been a stopgap solution, but this adds a subjective element to the process workflow. An ideal solution would wholly remove the human element and determine an appropriate MMD-based quality target value using solely the data provided. In this thesis, we present a solution for this situation, as we introduce a novel statistical test that can more accurately compare distributions of data and determine a target score using solely the sample sets. By inspecting the quality of this solution on various models, we can train and analyze models that perform at sufficiently good levels via a fully automated procedure.

In this thesis, we study different two-player zero-sum games, where one player, called Eve, has a reachability objective (i.e., aims to reach a particular configuration) and the other, called Adam, has a safety objective (i.e., aims to avoid the configuration). We study a general class of games, called Attacker-Defender games, where the computational environment can vary from as simple as the integer line to n-dimensional topological braids. Similarly, the moves themselves can be simple vector addition or linear transformations defined by matrices. The main computational problem is to decide whether Eve has a winning strategy to reach the target configuration from the initial configuration, or whether the dual holds, that is, whether Adam can ensure that the target is never reached. The notion of a winning strategy is widely used in game semantics and its existence means that the player can ensure that his or her winning conditions are met, regardless of the actions of the opponent. It general, games provide a powerful framework to model and analyse interactive processes with uncontrollable adversaries. We formulated several Attacker-Defender games played on different mathematical domains with different transformations (moves), and identified classes of games, where the checking for existence of a winning strategy is undecidable. In other classes, where the problem is decidable, we established their computational complexity. In the thesis, we investigate four classes of games where determining the winner is undecidable: word games, where the players' moves are words over a group alphabet together with integer weights or where the moves are pairs of words over group alphabets; matrix games on vectors, where players transform a three-dimensional vector by linear transformations defined by 3×3 integer matrices; braid games, where players braid and unbraid a given braid; and last, but not least, games played on two-dimensional Z-VAS, closing the gap between decidable and undecidable cases and answering an existing open problem of the field. We also identified decidable fragments, such as word games, where the moves are over a single group alphabet, games on one-dimensional Z-VASS. For word games, we provide an upper-bound of EXPTIME , while for games on Z-VASS, tight bounds of EXPTIME-complete or EXPSPACE-complete, depending on the state structure. We also investigate single-player systems such as polynomial iteration and identity problem in matrix semigroups. We show that the reachability problem for polynomial iteration is PSPACE-complete while the identity problem for the Heisenberg group is in PTIME for dimension three and in EXPTIME for higher dimensions.

This thesis defines the Broadcasting Automata model as an intuitive and complete method of distributed pattern formation, partitioning and distributed geometric computation. The system is examined within the context of Swarm Robotics whereby large numbers of minimally complex robots may be deployed in a variety of circumstances and settings with goals as diverse as from toxic spill containment to geological survey. Accomplishing these tasks with such simplistic machines is complex and has been deconstructed in to sub-problems considered to be signif- icant because, when composed, they are able to solve much more complex tasks. Sub-problems have been identified, and studied as pattern formation, leader elec- tion, aggregation, chain formation, hole avoidance, foraging, path formation, etc. The Broadcasting Automata draws inspiration from a variety of sources such as Ad-Hoc radio networks, cellular automata, neighbourhood sequences and nature, employing many of the same pattern forming methods that can be seen in the superposition of waves and resonance. To this end the thesis gives an in depth analysis of the primitive tools of the Broadcasting Automata model, nodal patterns, where waves from a variety of transmitters can in linear time construct partitions and patterns with results per- taining to the numbers of different patterns and partitions, along with the number of those that differ, are given. Using these primitives of the model a variety of algorithms are given including leader election, through the location of the centre of a discrete disc, and a solution to the Firing Squad Synchronisation problem. These problems are solved linearly.An exploration of the ability to vary the broadcasting radius of each node leads to results of categorisations of digital discs, their form, composition, encodings and generation. Results pertaining to the nodal patterns generated by arbitrary transmission radii on the plane are explored with a connection to broadcasting sequences and approximation of discrete metrics of which results are given for the approximation of astroids, a previously unachievable concave metric, through a novel application of the aggregation of waves via a number of explored functions. Broadcasting Automata aims to place itself as a robust and complete linear time and large scale system for the construction of patterns, partitions and geometric computation. Algorithms and methodologies are given for the solution of problems within Swarm Robotics and an extension to neighbourhood sequences. It is also hoped that it opens up a new area of research that can expand many older and more mature works.