1,721,083 research outputs found
Coevolution of Neuro-controllers to Train Multi-Agent Teams from Zero Knowledge
Get PDFDissertation (MSc)--University of Pretoria, 2013.After the historic chess match between Deep Blue and Garry Kasparov, many researchers considered the game of chess solved and moved on to the more complex game of soccer. Artificial intelligence research has shifted focus to creating artificial players capable of mimicking the task of playing soccer. A new training algorithm is presented in this thesis for training teams of players from zero knowledge, evaluated on a simplified version of the game of soccer. The new algorithm makes use of the charged particle swarm optimiser as a neural network trainer in a coevolutionary training environment. To counter the lack of domain information a new relative fitness measure based on the FIFA league-ranking system was developed. The function provides a granular relative performance measure for competitive training. Gameplay strategies that resulted from the trained players are evaluated. It was found that the algorithm successfully trains teams of agents to play in a cooperative manner. Techniques developed in this study may also be widely applied to various other artificial intelligence fields.Computer Scienceunrestricte
Derating NichePSO
Get PDFDissertation (MSc)--University of Pretoria, 2007.The search for multiple solutions is applicable to many fields (Engineering [54][67], Science [75][80][79][84][86], Economics [13][59], and others [51]). Multiple solutions allow for human judgement to select the best solution from a group of solutions that best match the search criteria. Finding multiple solutions to an optimisation problem has shown to be difficult to solve. Evolutionary computation (EC) and more recently Particle Swarm Optimisation (PSO) algorithms have been used in this field to locate and maintain multiple solutions with fair success. This thesis develops and empirically analyses a new method to find multiple solutions within a convoluted search space. The method is a hybrid of the NichePSO [14] and the sequential niche technique (SNT)[8]. The original SNT was developed using a Genetic Algorithm (GA). It included restrictions such as knowing or approximating the number of solutions that exist. A further pitfall of the SNT is that it introduces false optima after modifying the search space, thereby reducing the accuracy of the solutions. However, this can be resolved with a local search in the unmodified search space. Other sequential niching algorithms require that the search be repeated sequentially until all solutions are found without considering what was learned in previous iterations, resulting in a blind and wasteful search. The NichePSO has shown to be more accurate than GA based algorithms [14][15]. It does not require knowledge of the number of solutions in the search space prior to the search process. However, the NichePSO does not scale well for problems with many optima [16]. The method developed in this thesis, referred to as the derating NichePSO, combines SNT with the NichePSO. The main objective of the derating NichePSO is to eliminate the inaccuracy of SNT and to improve the scalability of the NichePSO. The derating NichePSO is compared to the NichePSO, deterministic crowding [23] and the original SNT using various multimodal functions. The performance of the derating NichePSO is analysed and it is shown that the derating NichePSO is more accurate than SNT and more scalable than the NichePSO.Computer ScienceMScUnrestricte
Training support vector machines with particle swarms
No full textDissertation (MSc)--University of Pretoria, 2007.Particle swarms can easily be used to optimize a function with a set of linear equality constraints, by restricting the swarm’s movement to the constrained search space. A “Linear Particle Swarm Optimiser” and “Converging Linear Particle Swarm Optimiser” is developed to optimize linear equality-constrained functions. It is shown that if the entire swarm of particles is initialized to consist of only feasible solutions, then the swarm can optimize the constrained objective function without ever again considering the set of constraints. The Converging Linear Particle Swarm Optimiser overcomes the Linear Particle Swarm Optimiser’s possibility of premature convergence. Training a Support Vector Machine requires solving a constrained quadratic programming problem, and the Converging Linear Particle Swarm Optimiser ideally fits the needs of an optimization method for Support Vector Machine training. Particle swarms are intuitive and easy to implement, and is presented as an alternative to current numeric Support Vector Machine training methods.Computer ScienceUnrestricte
Multi-guided particle swarm optimization : a multi-objective particle swarm optimizer
Get PDFThesis (PhD)--University of Pretoria, 2017.An exploratory analysis in low-dimensional objective space of the vector evaluated particle swarm optimization (VEPSO) algorithm is presented. A novel visualization technique is presented and applied to perform the exploratory analysis. The exploratory analysis together with a quantitative analysis revealed that the VEPSO algorithm continues to explore without exploiting the well-performing areas of the search space. A detailed investigation into the influence that the choice of archive implementation has on the performance of the VEPSO algorithm is presented. Both the Pareto-optimal front (POF) solution diversity and convergence towards the true POF is considered during the investigation. Attainment surfaces are investigated for their suitability in efficiently comparing two multi-objective optimization (MOO) algorithms. A new measure to objectively compare algorithms in multi-dimensional objective space, based on attainment surfaces, is presented. This measure, referred to as the porcupine measure, adapts the attainment surface measure by using a statistical test along with weighted intersection lines. Loosely based on the VEPSO algorithm, the multi-guided particle swarm optimization (MGPSO) algorithm is presented and evaluated. The results indicate that the MGPSO algorithm overcomes the weaknesses of the VEPSO algorithm and also outperforms a number of state of the art MOO algorithms on at least two benchmark test sets.Computer SciencePhDUnrestricte
Analysis of RED packet loss performance in a simulated IP WAN
Get PDFDissertation (MEng)--University of Pretoria, 2013.The Internet supports a diverse number of applications, which have different requirements for a number of services. Next generation networks provide high speed connectivity between hosts, which leaves the service provider to configure network devices appropriately, in order to maximize network performance. Service provider settings are based on best recommendation parameters, which give an opportunity to optimize these settings even further. This dissertation focuses on a packet discarding algorithm, known as random early detection (RED), to determine parameters which will maximize utilization of a resource. The two dominant traffic protocols used across an IP backbone are UDP and TCP. UDP traffic flows transmit packets regardless of network conditions, dropping packets without changing its transmission rates. However, TCP traffic flows concern itself with the network condition, reducing its packet transmission rate based on packet loss. Packet loss indicates that a network is congested. The sliding window concept, also known as the TCP congestion window, adjusts to the amount of acknowledgements the source node receives from the destination node. This paradigm provides a means to transmit data across the available bandwidth across a network. A well known and widely implemented simulation environment, the network simulator 2 (NS2), was used to analyze the RED mechanism. The network simulator 2 (NS2) software gained its popularity as being a complex networking simulation tool. Network protocol traffic (UDP and TCP) characteristics comply with theory, which verifies that the traffic generated by this simulator is valid. It is shown that the autocorrelation function differs between these two traffic types, verifying that the generated traffic does conform to theoretical and practical results. UDP traffic has a short-range dependency while TCP traffic has a long-range dependency. Simulation results show the effects of the RED algorithm on network traffic and equipment performance. It is shown that random packet discarding improves source transmission rate stabilization, as well as node utilization. If the packet dropping probability is set high, the TCP source transmission rates will be low, but a low packet drop probability provides high transmission rates to a few sources and low transmission rates to the majority of other sources. Therefore, an ideal packet drop probability was obtained to complement TCP source transmission rates and node utilization. Statistical distributions were fitted to sampled data from the simulations, which also show improvements to the network with random packet discarding. The results obtained contribute to congestion control across wide area networks. Even though a number of queuing management implementation exists, RED is the most widely used implementation used by service providers.Electrical, Electronic and Computer Engineeringunrestricte
Mining continuous classes using evolutionary computing
Get PDFDissertation (MSc)--University of Pretoria, 2006.Data mining is the term given to knowledge discovery paradigms that attempt to infer knowledge, in the form of rules, from structured data using machine learning algorithms. Specifically, data mining attempts to infer rules that are accurate, crisp, comprehensible and interesting. There are not many data mining algorithms for mining continuous classes. This thesis develops a new approach for mining continuous classes. The approach is based on a genetic program, which utilises an efficient genetic algorithm approach to evolve the non-linear regressions described by the leaf nodes of individuals in the genetic program's population. The approach also optimises the learning process by using an efficient, fast data clustering algo¬rithm to reduce the training pattern search space. Experimental results from both algorithms are compared with results obtained from a neural network. The experimental results of the genetic program is also compared against a commercial data mining package (Cubist). These results indicate that the genetic algorithm technique is substantially faster than the neural network, and produces comparable accuracy. The genetic program produces substantially less complex rules than that of both the neural network and Cubist.Computer Scienceunrestricte
Digital holographic microscopy with automated detection of red blood cells
Get PDFDissertation (MSc)--University of Pretoria, 2017.The digital in-line holographic configuration is motivated by the goal of developing a portable, cost effective sensor system for pre-screening patient blood samples. The theory of holography is explained from the foundational concepts in scalar diffraction theory all the way through to the implementation of reconstruction algorithms. Methods for the enhancement of holographic reconstructions are described. The algorithms that perform an automated count of the reconstructed objects are described and demonstrated. Simulated and experimental results are provided. Together, the lens-free holographic microscopy of micro-sized particles along with the application of image processing techniques for the automated detection and counting of objects of interest, provide a component towards realising a sensor system that can be used for pre-screening patient blood samples.CSIRComputer ScienceMScUnrestricte
Interactive narrative generation using computational verb theory
Get PDFDissertation (MSc)--University of Pretoria, 2009.Interactive narrative extends traditional story-telling techniques by enabling previously passive observers to become active participants in the narrative events that unfold. A variety of approaches have attempted to construct such interactive narrative spaces and reconcile the goals of interactivity and dramatic story-telling. With the advent of the linguistic variable in 1972, a means was established for modelling natural language words and phrases mathematically and computationally. Over the past decade, the computational verb, first introduced in 1997, has been developed as a mathematical means of modelling natural language verbs in terms of dynamic systems, and vice versa. Computational verb theory extends the initial concept of the linguistic variable beyond being able to model adjectives, nouns, and passive states, into the realm of actions as denoted by natural language verbs. This thesis presents the framework and implementation of a system that generates interactive narrative spaces from narrative text. The concept of interactive narrative is introduced and recent developments in the area of interactive narrative are discussed. Secondly, a brief history of the development of the linguistic variable and the computational verb are provided. With the context of the computational verb (interactive) narrative generation (CVTNG) system presented, the underlying theoretical principles of the system are established. The CVTNG system principles are described in terms of fuzzy set, computational verb, and constraint satisfaction theory. The fuzzy set, computational verb, and constraint satisfaction principles are organised according to a CVTNG architecture. The CVTNG architecture is then described in terms of its subsystems, structures, algorithms, and interfaces. Each CVTNG system component is related to the overall design considerations and goals. A prototype of the CVTNG system is implemented and tested against a suite of natural language sentences. The behaviour and performance of the CVTNG system prototype are discussed in relation to the CVTNG system’s design principles. Results are calculated and stored as variable values that are dynamically and generically associated with representational means, specifically computer graphics, to illustrate the generation of interactive narrative spaces. Plans for future work are discussed to show the immense development potential of this application. The thesis concludes that the CVTNG system provides a solid and extendable base for the intuitive generation of interactive narrative spaces from narrative text, computational verb models, and freely associated media. CopyrightComputer ScienceUnrestricte
Ant colony optimisation algorithms for solving multi-objective power-aware metrics for mobile ad hoc networks
Get PDFThesis (PhD)--University of Pretoria, 2010.A mobile ad hoc network (MANET) is an infrastructure-less multi-hop network where each node communicates with other nodes directly or indirectly through intermediate nodes. Thus, all nodes in a MANET basically function as mobile routers participating in some routing protocol required for deciding and maintaining the routes. Since MANETs are infrastructure-less, self-organizing, rapidly deployable wireless networks, they are highly suitable for applications such as military tactical operations, search and rescue missions, disaster relief operations, and target tracking. Building such ad-hoc networks poses a significant technical challenge because of energy constraints and specifically in relation to the application of wireless network protocols. As a result of its highly dynamic and distributed nature, the routing layer within the wireless network protocol stack, presents one of the key technical challenges in MANETs. In particular, energy efficient routing may be the most important design criterion for MANETs since mobile nodes are powered by batteries with limited capacity and variable recharge frequency, according to application demand. In order to conserve power it is essential that a routing protocol be designed to guarantee data delivery even should most of the nodes be asleep and not forwarding packets to other nodes. Load distribution constitutes another important approach to the optimisation of active communication energy. Load distribution enables the maximisation of the network lifetime by facilitating the avoidance of over-utilised nodes when a route is in the process of being selected. Routing algorithms for mobile networks that attempt to optimise routes while at- tempting to retain a small message overhead and maximise the network lifetime has been put forward. However certain of these routing protocols have proved to have a negative impact on node and network lives by inadvertently over-utilising the energy resources of a small set of nodes in favour of others. The conservation of power and careful sharing of the cost of routing packets would ensure an increase in both node and network lifetimes. This thesis proposes simultaneously, by using an ant colony optimisation (ACO) approach, to optimise five power-aware metrics that do result in energy-efficient routes and also to maximise the MANET's lifetime while taking into consideration a realistic mobility model. By using ACO algorithms a set of optimal solutions - the Pareto-optimal set - is found. This thesis proposes five algorithms to solve the multi-objective problem in the routing domain. The first two algorithms, namely, the energy e±ciency for a mobile network using a multi-objective, ant colony optimisation, multi-pheromone (EEMACOMP) algorithm and the energy efficiency for a mobile network using a multi-objective, ant colony optimisation, multi-heuristic (EEMACOMH) algorithm are both adaptations of multi-objective, ant colony optimisation algorithms (MOACO) which are based on the ant colony system (ACS) algorithm. The new algorithms are constructive which means that in every iteration, every ant builds a complete solution. In order to guide the transition from one state to another, the algorithms use pheromone and heuristic information. The next two algorithms, namely, the energy efficiency for a mobile network using a multi-objective, MAX-MIN ant system optimisation, multi-pheromone (EEMMASMP) algorithm and the energy efficiency for a mobile network using a multi-objective, MAX- MIN ant system optimisation, multi-heuristic (EEMMASMH) algorithm, both solve the above multi-objective problem by using an adaptation of the MAX-MIN ant system optimisation algorithm. The last algorithm implemented, namely, the energy efficiency for a mobile network using a multi-objective, ant colony optimisation, multi-colony (EEMACOMC) algorithm uses a multiple colony ACO algorithm. From the experimental results the final conclusions may be summarised as follows: Ant colony, multi-objective optimisation algorithms are suitable for mobile ad hoc networks. These algorithms allow for high adaptation to frequent changes in the topology of the network. All five algorithms yielded substantially better results than the non-dominated sorting genetic algorithm (NSGA-II) in terms of the quality of the solution. All the results prove that the EEMACOMP outperforms the other four ACO algorithms as well as the NSGA-II algorithm in terms of the number of solutions, closeness to the true Pareto front and diversity. Computer Scienceunrestricte
The feature detection rule and its application within the negative selection algorithm
Get PDFDissertation (MSc)--University of Pretoria, 2009.The negative selection algorithm developed by Forrest et al. was inspired by the manner in which T-cell lymphocytes mature within the thymus before being released into the blood system. The resultant T-cell lymphocytes, which are then released into the blood, exhibit an interesting characteristic: they are only activated by non-self cells that invade the human body. The work presented in this thesis examines the current body of research on the negative selection theory and introduces a new affinity threshold function, called the feature-detection rule. The feature-detection rule utilises the inter-relationship between both adjacent and non-adjacent features within a particular problem domain to determine if an artificial lymphocyte is activated by a particular antigen. The performance of the feature-detection rule is contrasted with traditional affinity-matching functions currently employed within negative selection theory, most notably the r-chunks rule (which subsumes the r-contiguous bits rule) and the hamming-distance rule. The performance will be characterised by considering the detection rate, false-alarm rate, degree of generalisation and degree of overfitting. The thesis will show that the feature-detection rule is superior to the r-chunks rule and the hamming-distance rule, in that the feature-detection rule requires a much smaller number of detectors to achieve greater detection rates and less false-alarm rates. The thesis additionally refutes that the way in which permutation masks are currently applied within negative selection theory is incorrect and counterproductive, while placing the feature-detection rule within the spectrum of affinity-matching functions currently employed by artificial immune-system (AIS) researchers.Computer ScienceUnrestricte
- …
