Terence Bossomaier

Phase-transition–like behaviour of information measures in financial markets

We apply measures based on information theory to the analysis of day close equity prices traded on US stock markets over the 13-year interval from 1995 up until after the market crash of September 2008. We show that the mutual information between prices provides insight into the changing relationships between equities over a time period which includes three known market crashes and two events which have not previously been included in this type of study, one of which is related to the sub-prime meltdown starting in 2007. Specifically, the mutual information around market crashes shows behaviour typical of the phase transitions studied in condensed-matter physics, however similar but more extended peaks in mutual information are also observed at other times not associated with any known market crashes.

Information and phase transitions in socio-economic systems

We examine the role of information-based measures in detecting and analysing phase transitions. We contend that phase transitions have a general character, visible in transitions in systems as diverse as classical flocking models, human expertise, and social networks. Information-based measures such as mutual information and transfer entropy are particularly suited to detecting the change in scale and range of coupling in systems that herald a phase transition in progress, but their use is not necessarily straightforward, possessing difficulties in accurate estimation due to limited sample sizes and the complexities of analysing non-stationary time series. These difficulties are surmountable with careful experimental choices. Their effectiveness in revealing unexpected connections between diverse systems makes them a promising tool for future research.

Agent Heterogeneity and Coalition Formation: Investigating Market-Based Cooperative Problem Solving

One focus of multi-agent systems research is the notion that complex outcomes or behaviours may be arrived at through the interaction of agents. However, it is still an open question as to how agents in a complex system form coalitions or modules, and how these coalitions self-organize into hierarchies. In this paper, we begin to address this question by investigating agent collaboration in the context of a high-level pattern recognition task. We propose a novel market-based communication protocol, which governs the aggregate behaviour of individual agents and subsequent emergent properties of the system. Based on the Contract Net Protocol, individual agents bid to join coalitions (or solutions to a given problem). An important contribution of this study is the analysis of the role heterogeneous agents play in the formation of coalitions. Using a simple model, we show that by promoting diversity within the agent population it is possible to avoid deadlock or tie conditions, which otherwise have to be solved arbitrarily by the deadlocked agents.

The Perceptual Cues that Reshape Expert Reasoning

The earliest stages in our perception of the world have a subtle but powerful influence on later thought processes; they provide the contextual cues within which our thoughts are framed and they adapt to many different environments throughout our lives. Understanding the changes in these cues is crucial to understanding how our perceptual ability develops, but these changes are often difficult to quantify in sufficiently complex tasks where objective measures of development are available. Here we simulate perceptual learning using neural networks and demonstrate fundamental changes in these cues as a function of skill. These cues are cognitively grouped together to form perceptual templates that enable rapid ‘whole scene categorisation of complex stimuli. Such categories reduce the computational load on our capacity limited thought processes, they inform our higher cognitive processes and they suggest a framework of perceptual pre-processing that captures the central role of perception in expertise.

Social media in Ebola outbreak.

The West African 2014 Ebola outbreak has highlighted the need for a better information network. Hybrid information networks, an integration of both hierarchical and formalized command control-driven and community-based, or ad hoc emerging networks, could assist in improving public health responses. By filling the missing gaps with social media use, the public health response could be more proactive rather than reactive in responding to such an outbreak of global concern. This article provides a review of the current social media use specifically in this outbreak by systematically collecting data from ProQuest Newsstand, Dow Jones Factiva, Program for Monitoring Emerging Diseases (ProMED) as well as Google Trends. The period studied is from 19 March 2014 (first request for information on ProMED) to 15 October 2014, a total of 31 weeks. The term Ebola was used in the search for media reports. The outcome of the review shows positive results for social media use in effective surveillance response mechanisms - for improving the detection, preparedness and response of the outbreak - as a complement to traditional, filed, work-based surveillance approach.

DenClust: A Density Based Seed Selection Approach for K-Means

In this paper we present a clustering technique called DenClust that produces high quality initial seeds through a deterministic process without requiring an user input on the number of clusters k and the radius of the clusters r. The high quality seeds are given input to K-Means as the set of initial seeds to produce the final clusters. DenClust uses a density based approach for initial seed selection. It calculates the density of each record, where the density of a record is the number of records that have the minimum distances with the record. This approach is expected to produce high quality initial seeds for K-Means resulting in high quality clusters from a dataset. The performance of DenClust is compared with five (5) existing techniques namely CRUDAW, AGCUK, Simple K-means (SK), Basic Farthest Point Heuristic (BFPH) and New Farthest Point Heuristic (NFPH) in terms of three (3) external cluster evaluation criteria namely F-Measure, Entropy, Purity and two (2) internal cluster evaluation criteria namely Xie-Beni Index (XB) and Sum of Square Error (SSE). We use three (3) natural datasets that we obtain from the UCI machine learning repository. DenClust performs better than all five existing techniques in terms of all five evaluation criteria for all three datasets used in this study.

Neuro-cognitive model of move location in the game of Go

Although computer Go players are now better than humans on small board sizes, they are still a fair way from the top human players on standard board sizes. Thus the nature of human expertise is of great interest to artificial intelligence. Human play relies much more on pattern memory and has been extensively explored in chess. The big challenge in Go is local-global interaction - local search is good but global integration is weak. We used techniques based on the cognitive neuroscience of chess to predict optimal areas to move using perceptual chunks, which we cross-validated against game records comprising upwards of five million positions. Prediction to within a small window was about 50%, a remarkable result.

Spoken communication with computer game characters

This paper explores the characteristics of a new language for communication with game characters. This new language is called game pidgin language (GPL) (Rudra et al., 2003). A GPL is a sub-class of computer pidgin language (CPL) (Hinde and Belrose) which is a new spoken language taught to the game player and is efficient for dialogues with the computer. The GPL is extended by adding three conditions to the grammar. We add some attributes in an already established eXtensible Markup Language (XML) to encapsulate a cognitive response from the non-player-character (NPC). A complete game pidgin model is proposed in this paper.

Towards realtime stance classification by spiking neural network

Spiking neural networks are a popular area of current research in both artificial intelligence and neuroscience. Unlike second generation networks like the multilayer perceptron (MLP), they simulate rather than emulate neuronal interactions. Spiking networks have been shown to be theoretically more powerful than earlier generation networks, and have repeatedly been suggested as ideal for realtime problems due to their time-basis. Because of their sparse nature, real neural networks are also extremely power-efficient, a pressing concern in computing today. This raises the possibility of applying sparse spiking networks for power-saving. To investigate these ideas, we wish to apply a spiking network to realtime data classification. As a first step, we use a feedforward network with the SpikeProp algorithm to classify offline skeleton data derived from a depth camera. Classifier networks were successfully trained, but we found SpikeProp considerably more complex to apply than backpropagation. There is considerable potential for optimization and power efficiency, and we hope to compare the performance of our system with more established learning techniques in a realtime setting.

Cognitive Emotion in Speech Interactive Games

This paper falls into the category of human computer interaction. Here the focus lies in injecting emotions into the non-player characters (NPC) in reaction to the emotion of the player in computer games. The idea of injecting emotion into games has been a popular topic for discussion in the games industry today. The paper reviews the emotional side of games from both the game side and the human player side, considering speech style and content, facial expressions and gestures. The emotional content is encapsulated within an XML and an associated finite state automaton.