Novel Integrated Methods of Evolving Spiking Neural Network and Particle Swarm Optimisation
| aut.embargo | No | en_NZ |
| aut.thirdpc.contains | No | en_NZ |
| aut.thirdpc.permission | No | en_NZ |
| aut.thirdpc.removed | No | en_NZ |
| dc.contributor.advisor | Kasabov, Nikola | |
| dc.contributor.advisor | Shamsuddin, Siti Mariyam | |
| dc.contributor.advisor | Schliebs, Stefan | |
| dc.contributor.author | Abdull Hamed, Haza Nuzly | |
| dc.date.accessioned | 2012-06-18T00:09:47Z | |
| dc.date.available | 2012-06-18T00:09:47Z | |
| dc.date.copyright | 2012 | |
| dc.date.created | 2012 | |
| dc.date.issued | 2012 | |
| dc.date.updated | 2012-06-17T02:42:29Z | |
| dc.description.abstract | This thesis proposes and presents several methods for classification problems. Spatial and spatiotemporal classification problems have been considered in this study. A novel integration between Evolving Spiking Neural Network (ESNN) and Particle Swarm Optimisation (PSO) is proposed for ESNN model optimisation. ESNN, motivated by the principle of Evolving Connectionist System (ECoS), is a relatively new classifier in the neural information processing area. Proper combination of ESNN parameters would influence the ESNN performance. On the other hand, PSO is a bio-inspired optimiser and was developed based on a study of school of fish and flock of birds behaviour. In this framework, all ESNN parameters are optimised by the PSO to achieve optimal parameter combination for the model. A wrapper approach is implemented in the ESNN-PSO frameworks and a few other integrated frameworks that are also proposed in this work. The classifier uses information provided by the particles during learning and generates a fitness value for each solution candidate. Particles interact with each other and update their information based on the global best particle (gbest) and their own best solution (pbest). The learning process continues until termination criteria are met. When dealing with high dimensional problems, only some of the input features are relevant. In this case, selection of features is required. Since standard PSO is not able to handle probability computation, the quantum computation principle is embedded into PSO. This combination is referred to as Quantum-inspired Particle Swarm Optimisation (QiPSO). The integrated ESNN-QiPSO is proposed in this study for simultaneous feature selection and parameter optimisation. This combination provides promising results that may lead to better and faster learning. However, several problems have been identified that led to the development of enhanced QiPSO and ESNN. A hybrid particle and new search and update strategy is proposed for the QiPSO and is presented in the Dynamic QiPSO (DQiPSO) model. Subsequently, an integrated framework of DQiPSO and ESNN is proposed for efficient feature selection and parameter optimisation. The probabilistic element is also embedded into ESNN as part of its enhancement. In the Probabilistic ESNN (PESNN), the evolving connection is introduced. In the proposed integrated PESNN-DQiPSO, the classifier works together with the optimiser where the connection, feature and parameter components are optimised synchronously for better classification. Real world problems are often spatiotemporal. Standard ESNN architecture lacks the ability to process both spatial and temporal components in spatiotemporal problems. This study proposes two new ESNN frameworks for spatiotemporal classification utilising the reservoir computing principle. The Extended ESNN (EESNN) is proposed where a simple memory is used to accumulate all spatial and temporal information before passing them to ESNN. In the second approach, more complex Liquid State Machine (LSM) reservoir is incorporated into the ESNN. The reservoir-based ESNN (RESNN) accumulates all information and generates the reservoir responses that can be measured at any simulation time. These responses are encoded into liquid states before sending them to ESNN for classification. All proposed frameworks have been evaluated on synthetic and real world problems. This study also proposes a spatiotemporal synthetic problem called Rotating Dot. The purpose of introducing this benchmark dataset is to have a spatiotemporal problem with controllable difficulty that can be used for evaluation of the methods. In this problem, the noise can be set at a small value to generate a simple problem or at a high value for more difficult problems. Results obtained with all proposed frameworks are promising and warrant future exploration. | en_NZ |
| dc.identifier.uri | https://hdl.handle.net/10292/4459 | |
| dc.language.iso | en | en_NZ |
| dc.publisher | Auckland University of Technology | |
| dc.rights.accessrights | OpenAccess | |
| dc.subject | Spiking Neural Network | en_NZ |
| dc.subject | Particle Swarm Optimisation | en_NZ |
| dc.subject | Evolving Connectionist System | en_NZ |
| dc.subject | Quantum Computation | en_NZ |
| dc.subject | Liquid State Machine | en_NZ |
| dc.subject | Classification | en_NZ |
| dc.subject | Spatiotemporal | en_NZ |
| dc.subject | Spatial | en_NZ |
| dc.subject | Feature Selection | en_NZ |
| dc.subject | Parameter Optimisation | en_NZ |
| dc.title | Novel Integrated Methods of Evolving Spiking Neural Network and Particle Swarm Optimisation | en_NZ |
| dc.type | Thesis | |
| thesis.degree.discipline | ||
| thesis.degree.grantor | Auckland University of Technology | |
| thesis.degree.level | Doctoral Theses | |
| thesis.degree.name | Doctor of Philosophy | en_NZ |