Evolutionary robotics controllers with location perception facilitated by neural network-based simulators
- Authors: Phillips, Antin Paul
- Date: 2021-04
- Subjects: Evolutionary Robots -- South Africa , Neural networks (Computer science) -- South Africa
- Language: English
- Type: Master's theses , text
- Identifier: http://hdl.handle.net/10948/61613 , vital:71474
- Description: Humans impressively maintain a real-time approximation of their bodily form. For instance, one knows where one’s arm is, relative to the body, without needing to directly observe it. This ability, in part, allows humans to interact with the environment without direct observation. This bodily sense is referred to as ”proprioception“. The human body contains various proprioceptors, sensory neurons which provide information about the physical state of the body. This information, along with internal body representations that humans develop over time, allows one to maintain an approximation of their bodily form. Humans also possess an impressive sense of direction and navigation ability. For instance, a blindfolded human can move around a familiar environment and maintain an approximate sense of where they are within that environment. This ability is, in part, enabled by proprioception as it provides one with an approximation of the effects their actions have on their body. The field of Evolutionary Robots (ER) makes extensive use of robotic simulators to carry out simulated robotic evaluations. Research has been conducted into alternate forms of simulation and Simulator Neural Networks (SNNs) were subsequently developed. The speed and accuracy of these SNNs, relative to more typical simulation techniques, is what inspired the approach explored in this research. Robots do not necessarily possess the appropriate hardware to sense their position within an environment. Thus, it was proposed that SNNs could be incorporated into ER controllers to approximate the position of the robot. These SNNs would be executed in parallel to the robot and provide a constant approximation of the robot’s position. This would provide controllers with information that they would not otherwise have, albeit approximate information. Various experiments were carried out which examined both typical ER controllers as well as those which were augmented in the proposed fashion. The augmented controllers were found to outperform typical controllers as well as develop more advanced and efficient behaviours. Furthermore, the augmented controllers demonstrated the ability to solve tasks that regular controllers could not. A potential criticism of the approach suggested in this research is that ER controllers could hypothetically be trained in such a way that the proposed augmentation would be unnecessary. This possibility was investigated and it was found that successfully training controllers in such a manner would be unlikely. Furthermore, the effort involved in fine-tuning this training process would be greater than simply following the approach suggested in this research. Another potential drawback of the suggested approach involved the accuracy of the information that SNNs could provide to controllers. The approximated information was found to diverge over time and negatively affected controller performance. A method to address this issue was proposed and subsequently implemented. This method was demonstrated to be an effective means of reducing the divergence of the SNNs outputs and, in turn, improved controller performance. , Thesis (MSc) -- Faculty of Science, School of Computer Science, Mathematics, Physics and Statistics, 2021
- Full Text:
- Date Issued: 2021-04
Evolutionary robotics controllers with location perception facilitated by neural network-based simulators
- Authors: Phillips, Antin Paul
- Date: 2021-04
- Subjects: Grahamstown (South Africa) , Eastern Cape (South Africa) , Neural networks (Computer science) -- South Africa
- Language: English
- Type: Master's theses , text
- Identifier: http://hdl.handle.net/10948/52137 , vital:43444
- Description: Humans impressively maintain a real-time approximation of their bodily form. For instance, one knows where one’s arm is, relative to the body, without needing to directly observe it. This ability, in part, allows humans to interact with the environment without direct observation. This bodily sense is referred to as ”proprioception“. The human body contains various proprioceptors, sensory neurons which provide information about the physical state of the body. This information, along with internal body representations that humans develop over time, allows one to maintain an approximation of their bodily form. Humans also possess an impressive sense of direction and navigation ability. For instance, a blindfolded human can move around a familiar environment and maintain an approximate sense of where they are within that environment. This ability is, in part, enabled by proprioception as it provides one with an approximation of the effects their actions have on their body. The field of Evolutionary Robots (ER) makes extensive use of robotic simulators to carry out simulated robotic evaluations. Research has been conducted into alternate forms of simulation and Simulator Neural Networks (SNNs) were subsequently developed. The speed and accuracy of these SNNs, relative to more typical simulation techniques, is what inspired the approach explored in this research. Robots do not necessarily possess the appropriate hardware to sense their position within an environment. Thus, it was proposed that SNNs could be incorporated into ER controllers to approximate the position of the robot. These SNNs would be executed in parallel to the robot and provide a constant approximation of the robot’s position. This would provide controllers with information that they would not otherwise have, albeit approximate information. Various experiments were carried out which examined both typical ER controllers as well as those which were augmented in the proposed fashion. The augmented controllers were found to outperform typical controllers as well as develop more advanced and efficient behaviours. Furthermore, the augmented controllers demonstrated the ability to solve tasks that regular controllers could not. A potential criticism of the approach suggested in this research is that ER controllers could hypothetically be trained in such a way that the proposed augmentation would be unnecessary. This possibility was investigated and it was found that successfully training controllers in such a manner would be unlikely. Furthermore, the effort involved in fine-tuning this training process would be greater than simply following the approach suggested in this research. Another potential drawback of the suggested approach involved the accuracy of the information that SNNs could provide to controllers. The approximated information was found to diverge over time and negatively affected controller performance. A method to address this issue was proposed and subsequently implemented. This method was demonstrated to be an effective means of reducing the divergence of the SNNs outputs and, in turn, improved controller performance. , Thesis (MSc) -- Faculty of Science, School of Computer Science, Mathematics, Physics and Statistics, 2021
- Full Text: false
- Date Issued: 2021-04