Browsing by Author "Nitschke, Geoff"
Now showing 1 - 5 of 5
Results Per Page
Sort Options
- ItemOpen AccessDistributed autonomous intersection management with neuro-evolution(2021) Cherry, Matthew P; Nitschke, GeoffThe sudden surge in computational power available to computer research in industry and academia has led to developments in AI automation. More and more tasks are able to be automated and replaced with machine learning systems. One such task that promises to be highly beneficial is that of driving, clearly indicated by the amount of resources being spent by companies such as Uber, Google and Tesla. Neuro-Evolution has shown promise in the field of controller development, due to its ability to develop complex behaviour without a need for any labelled training data. It has been applied previously in car controller generation, across many fields. This thesis aims to apply Neuro-Evolution specifically to the field of intersection management, in order to study which methods are the most effective for this particular task. In particular we investigate three key hyper-parameters: Neuro-Evolution algorithm, task difficulty and problem exposure. A traffic simulator was developed and the hyper-parameters were used to evolve car controllers, which where then tested on unseen tasks. We show that certain key combinations of hyper-parameters yield exceptional results, but that direct correlations between individual parameters and performance are unclear, indicating that these methods are highly sensitive to hyper-parameter selection. We further identify some areas in which to optimize the evolution method, by looking at hyper-parameters which have a computational cost but which did not produce better performance.
- ItemOpen AccessMulti-objective evolutionary algorithms for product design(2024) Aslan, Bilal Hasan; Nitschke, GeoffIdentifying chemical compounds with optimal properties for specific applications presents a fundamental challenge in materials science. Traditional methods, based on trialand-error, are inefficient and costly. This thesis introduces an innovative integration of Computational Chemistry and Machine Learning (ML) with Evolutionary MultiObjective Optimisation (EMOO) techniques to streamline compound design. This approach automates the design process by leveraging ML to accurately predict compound properties and using EMOO to select compounds that meet various criteria. The significance of this work lies in its potential to transform the traditional development process, facilitating the creation of chemical products that fulfill multiple objectives more efficiently. This study not only demonstrates the synergy between advanced ML and optimisation techniques but also presents a comprehensive comparison of the MultiObjective Covariance Matrix Adaptation Evolution Strategy (MO-CMA-ES) and Nondominated Sorting Genetic Algorithm II (NSGA-II), including two novel meta-heuristics for enhanced molecular exploration. Our findings reveal that MO-CMA-ES, especially when combined with an extended search meta-heuristic, excels in exploring molecular spaces, establishing it as a preferred method for compound synthesis. This research promises to accelerate compound development specifically for detergent compounds, offering significant implications for product design across various industries.
- ItemOpen AccessNeuro-evolution search methodologies for collective self-driving vehicles(2019) Huang, Chien-Lun Allen; Nitschke, GeoffRecently there has been an increasing amount of research into autonomous vehicles for real-world driving. Much progress has been made in the past decade with many automotive manufacturers demonstrating real-world prototypes. Current predictions indicate that roads designed exclusively for autonomous vehicles will be constructed and thus this thesis explores the use of methods to automatically produce controllers for autonomous vehicles that must navigate with each other on these roads. Neuro-Evolution, a method that combines evolutionary algorithms with neural networks, has shown to be effective in reinforcement-learning, multi-agent tasks such as maze navigation, biped locomotion, autonomous racing vehicles and fin-less rocket control. Hence, a neuro-evolution method is selected and investigated for the controller evolution of collective autonomous vehicles in homogeneous teams. The impact of objective and non-objective search (and a combination of both, a hybrid method) for controller evolution is comparatively evaluated for robustness on a range of driving tasks and collection sizes. Results indicate that the objective search was able to generalise the best on unseen task environments compared to all other methods and the hybrid approach was able to yield desired task performance on evolution far earlier than both approaches but was unable to generalise as effectively over new environments.
- ItemOpen AccessPredicting diarrhoea outbreak with climate change(2021) Abdullahi, Tassallah Amina; Nitschke, GeoffClimate change is expected to exacerbate diarrhoea outbreak in South Africa, a leading cause of morbidity and mortality in the region. In this study, we modelled the impacts of climate change on diarrhoea with machine learning methods. We applied two deep learning techniques, convolutional neural networks (CNNs) and long-short term memory networks (LSTMs); and a support vector machine to predict daily diarrhoea cases over the different South African provinces by incorporating climate information. Generative Adversarial Networks (GANs) was used to generate synthetic data which was used to augment the available dataset. Furthermore, relevance estimation and value calibration (REVAC) was used to tune the parameters of the machine learning algorithms to optimize the accuracy of their predictions. Sensitivity analysis was also performed to investigate the contribution of the different climate factors to the diarrhoea prediction model. The results of the study showed that all three ML methods were appropriate for predicting daily diarrhoea cases with respect to the selected climate variables in each South African province. The ML methods were all able to yield low and similar RMSE. However, the level of accuracy for each model varied across different experiments, with the deep learning models outperforming the SVM model. Among the deep learning techniques, the CNN model performed best when only real-world dataset was used, while the LSTM model outperformed the other models when the real dataset was augmented with synthetic data. Across the provinces, the accuracy of all three ML algorithms improved by at least 30% when data augmentation was implemented. In addition, REVAC improved the accuracy of the CNN model by more than 12% in KwaZulu Natal province. However, the percentage increase in accuracy of the LSTM model was less than 4% in Western Cape province when REVAC was used. Our sensitivity analysis revealed that the most influential climate variables to be considered when predicting outbreak of diarrhoea in South Africa are precipitation, humidity, evaporation and temperature conditions. The result of this study is important for the development of an early warning system for diarrhoea outbreak over South Africa.
- ItemOpen AccessScalable hierarchical evolution strategies(2022) Abramowitz, Sasha; Nitschke, GeoffHierarchical reinforcement learning (HRL) has been steadily growing in popularity for solving the hardest reinforcement learning problems. However, current HRL algorithms are relatively slow and brittle to hyperparameter changes. This paper offers a solution to these slow and brittle HRL algorithms, by investigating a novel method combining Scalable Evolution Strategies (SES) and HRL. S-ES, named for its excellent scalability, was popularised by Open AI when they showed its performance to be comparable to state-of-the art policy gradient methods. However, S-ES has not been tested in conjunction with HRL methods, which empower temporal abstraction thus allowing agents to tackle more challenging problems. We introduce a novel method merging S-ES and HRL, which creates a highly scalable and fast (wall-clock time) algorithm. We demonstrate that S-ES needs no hyper-parameter tuning for the HRL tasks tested and is indifferent to delayed rewards. This results in a method that is significantly faster than gradient-based HRL methods while having competitive task performance. We extend this method using transfer learning with the aim of increasing task performance and novelty search with the goal of improving its exploration characteristics. The paper's main contribution is thus a novel evolutionary HRL method, namely Scalable Hierarchical Evolution Strategies, which yields greater learning speed and competitive task-performance compared to state-of-the-art gradient-based methods, across a range of tasks.