Spiking neural network for targeted navigation and collision avoidance in an autonomous robot
| dc.contributor.author | Ramne, Malin | |
| dc.contributor.department | Chalmers tekniska högskola / Institutionen för mekanik och maritima vetenskaper | sv |
| dc.contributor.examiner | Wahde, Mattias | |
| dc.contributor.supervisor | Wahde, Mattias | |
| dc.date.accessioned | 2021-09-17T09:00:55Z | |
| dc.date.available | 2021-09-17T09:00:55Z | |
| dc.date.issued | 2020 | sv |
| dc.date.submitted | 2020 | |
| dc.description.abstract | In this project a two-layered spiking neural network is implemented for targeted navigation and collision avoidance in an autonomous robot. The performance can be further improved by training with RM-STDP. The proposed modified STDP for inhibitory synapses yields even better performance in the networks and often after fewer rounds of training. The project also presents an evolutionary algorithm for determining the synaptic connections of a spiking neural network with a hidden layer. The evolutionary algorithm shows potential of working as a tool for determining the synaptic connectivity of spiking neural networks, however this project only explores a first rather simple implementation of the algorithm. In one run of the algorithm, with networks with a hidden layer consisting of 100 hidden neurons, a network with the ability to arrive at most of the test cases was evolved in less then 20 generations. However further work is necessary in order to determine the true potential of this approach. The performance of the spiking neural networks in this project are compared with the performance of non-spiking neural networks in order to determine advantages and disadvantages of using spiking neural networks in this specific case. The non-spiking neural networks perform better on the target navigation and collision avoidance tasks than corresponding spiking neural networks. The spiking neural network with no hidden layer arrived at the target in 842 out of 1000 test cases, and the non-spiking network arrived in 914 cases. With the hidden layer the number or arrivals where 784 for the spiking network and 916 for the non-spiking. This indicates that the advantages of the spike-rate interpretation of non-spiking neurons outweigh the expected advantages of encoding information in the specific timing of spikes in spiking neural networks, at least in the network structures and the tasks examined in this project. | sv |
| dc.identifier.coursecode | MMSX30 | sv |
| dc.identifier.uri | https://hdl.handle.net/20.500.12380/304140 | |
| dc.language.iso | eng | sv |
| dc.relation.ispartofseries | 2020:36 | sv |
| dc.setspec.uppsok | Technology | |
| dc.subject | spiking neural network | sv |
| dc.subject | spike timing dependent plasticity | sv |
| dc.subject | reward modulation | sv |
| dc.subject | evolutionary algorithm | sv |
| dc.title | Spiking neural network for targeted navigation and collision avoidance in an autonomous robot | sv |
| dc.type.degree | Examensarbete för masterexamen | sv |
| dc.type.uppsok | H | |
| local.programme | Complex adaptive systems (MPCAS), MSc |