Examensarbeten för masterexamen // Master Theses
Länka till denna samling:
Browse
Browsar Examensarbeten för masterexamen // Master Theses efter Författare "Andersson, Adam"
Visar 1 - 10 av 10
Sökresultat per sida
Sortera efter
- PostAdaptive Radar Illuminations with Deep Reinforcement Learning: Illumination Scheduling for Long Range Surveillance Radar with the use of Proximal Policy Optimization(2023) Sandelius, Samuel; Ekelund Karlsson, Albin; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Helgesson, Peter; Andersson, AdamA modern radar antenna can direct its energy electronically without inertia or the need for mechanically steering. This opens up several degrees of freedom such as transmission direction and illumination time, and thus also the potential to optimise operation in real-time. Long range surveillance radars solve the trade-off between searching for new targets and tracking known targets. This optimisation is often rule-based. In recent years, Reinforcement Learning (RL) Algorithms have been able to efficiently solve increasingly difficult tasks, such as mastering game strategies or solving complex control tasks. In this thesis we show that reinforcement learning can outperform such rule-based approaches for a simulated radar.
- PostDeep-learning-accelerated Bayesian inference for state-space models(2020) Hölén Hannouch, Elias; Holmstedt, Oskar; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Schauer, Moritz; Picchini, Umberto; Andersson, AdamBayesian inference is an important statistical tool for estimating uncertainties in model parameters from data. One very important method is the Metropolis-Hastings algorithm, which allows for parameter inference when analytical solutions are intractable. The only requirement is that the likelihood function can be evaluated. However, it is a computationally expensive algorithm, as it is usually run for several thousand iterations. This is especially true for inference in state-space models, where the likelihood is computed via Bayesian filtering, which is a costly operation in and of itself. We propose a new method for doing Bayesian inference via Metropolis- Hastings for state-space models by replacing the standard likelihood computation with a neural network. The network is trained on data generated by a much shorter earlier run of Metropolis-Hastings. We show, both qualitatively and quantitatively, that our method produces comparable results to the traditional method for several models. Moreover, our results indicate that the performance of our method is consistent as the dimensionality of the state-space model increases. Finally, we show that our method is much more computationally efficient than the traditional method for large runs. We investigate at what point our method becomes the preferable alternative and find that the threshold occurs at quite small runs, both in terms of computational time and desired output size.
- PostEstimation of Lidar Point Clouds Based on Ultrasonic Sensors Using Deep Learning(2023) Hjalmarsson, Carl; Jäghagen, Jesper; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Andersson, Adam; Abramsson, AndreasSafety is a key point of research in the automotive industry. Car companies dedicate a great amount of time and resources to make their cars safer by developing sensory systems like Low Speed Perception (LSP). In this thesis, we have explored the possibility to enhance the contribution of ultrasonic sensors to LSP by leveraging deep learning to mimic data from the more expensive Lidar sensors. To do this we draw inspiration from three different deep learning approaches: image denoising, image segmentation and image-to-image translation, resulting in five different models: DAE(bin), DAE(mse), UNET(bin), UNET(ce) and an I2I cGAN. We train these models on three datasets, each containing paired ultrasonic and Lidar representations of a two-dimensional environment around the car. In order to measure the performance of these models, we develop an evaluation framework where we assess the ability of the models to map ultrasonic object detections to corresponding Lidar detections. We find that the performance of all networks is highly dependent on the data representation. When using a basic representation, consisting only of point detections and free-space, all models fail to improve upon the baseline ultrasonic sensor score. When using a sparse representation consisting of detection arcs, however, UNET(bin) succeeds in outperforming the baseline and mimicking the more accurate Lidar representation (see cover). Finally, when adding unknown areas of the environment to the sparse representation, both UNET(ce) and the cGAN manage to outperform the baseline in most aspects, and we see a convergence towards the more realistic Lidar representation. The results show that there is indeed a possibility to enhance ultrasonic sensor perception using deep learning and Lidar reference data, and while there is still much room for improvement, we have shown that there is potential in further research on this task.
- PostExploration of Reinforcement Learning in Radar Scheduling(2021) Nathanson, Axel; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Larsson, Stig; Andersson, AdamThe development of phased array antennas has enhanced the effectiveness of radars thanks to it’s flexibility allowing the radar beam to be controlled and adapted almost instantaneously. This flexibility allows a radar to carry out multiple tasks simultaneously, such as surveillance of an area and tracking of targets. Traditionally the scheduling is performed according to hard-coded priority lists in combination with local optimisation, rather than global mathematical optimisation. Reinforcement learning algorithms have in the last few years successfully solved several artificial control tasks and is slowly starting to show some successes in realworld scenarios. Encouraged by the success we study the application of the Proximal Policy Optimisation (PPO) algorithm on a radar scheduling task. The algorithm is trained to track targets and search for new ones within a surveillance area. The proposed algorithm did not solve the scheduling task, but we identify and formalise the challenges that need to be addressed to be able to solve the radar scheduling task with the PPO algorithm.
- PostFast Bayesian Inference with Piecewise Deterministic Markov Processes(2023) Hammar, Karl; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Schauer, Moritz; Andersson, Adam; Svedung Wettervik, BenjaminPiecewise Deterministic Markov Processes (PDMPs) present a recent class of samplers for Bayesian inference. In this thesis, PDMP samplers are employed to sample state and latent parameters of an adversarial missile, described by an SDE. An approximate method for fast sampling is developed for this problem, and the performance of two different PDMP samplers, the Zig-Zag sampler, and the bouncy particle sampler, are compared. We find that the approximations needed for the methods to be competitive have a small impact on accuracy and that the method has the potential to be useful in real-world applications. Additionally, an approach for sampling from target models which may experience discontinuous jumps is developed. Using a particular trajectory realization of one such model we show that the method works as expected. This bears importance for the sampling of parameters related to maneuvering target types, where jump dynamics are relevant for target modeling.
- PostMachine learning algorithm for detecting periodic disturbances of microwave signals(2022) Lundmark, Oliver; Abelson, Julius; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Andersson, Adam; Sjödin, Martin; Tavara, ShirinAs digitalisation is expanding into new fields the demand for secure and stable connections between devices are ever-increasing. This can be done via microwave link networks, that comes at a fairly cheap price in comparison with fiber optic cables, but with the con of being more exposed to external disturbances. These disturbances could be caused by several different phenomena, including rain, wind and construction cranes. The scope of this thesis was to expand an already existing tool for detecting disturbances in Ericsson’s customer’s microwave link network, adding the possibility of detecting and classifying one more disturbance. This disturbance is caused by sunrays, which leads to a thermal expansion on one side of the mast where the node of a link is attached to, causing a miss-alignment between the antennas. The disturbance is called periodic sway, due to its characteristic 24 hours periodicity, correlating to the sun’s periodicity. The tool uses convolutional neural networks (CNN) to detect and classify disturbances. The CNN model needs features to train on to properly classify the disturbances. Today Ericsson’s tool uses the features like received signal power and attenuation. When expanding the tool and adding the periodic sway disturbance, further features had to be added, to capture the periodic nature of this particular disturbance. This resulted in adding historical data as a feature. The conclusions drawn from this thesis are that adding three days of historical data is sufficient for detecting and classifying this disturbance. Furthermore, the results imply that a sparse sampling period of 30-60 minutes is enough for the CNN to detect the periodicity.
- PostModelling of Human Behaviour in Traffic Interactions Using Inverse Reinforcement Learning(2021) Amaya, Santiago; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Andersson, Adam; Tarankanov, Yury
- PostOn nonlinear machine learning methodology for dose-response data in drug discovery(2020) Granbom, Klara; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Särkää, Aila; Andersson, AdamThis thesis investigates novel approaches to use nonlinear methodology for doseresponse data in drug discovery. Such methodology could potentially create insights and value within the field, saving resources such as time and usage of animals in experiments. Methods for dimensionality reduction and visualization, as well as methods for classification of compounds into clinical classes based on therapeutic usage, are investigated. The thesis builds partly upon previous research where linear methods, based on partial least squares and principal component analysis, have been used for dimensionality reduction in drug discovery. By using results from linear methods as a benchmark, this thesis investigates the nonlinear methods kernel partial least squares and t-distributed stochastic neighbor embedding for dimensionality reduction. Moreover, methods for classification of compounds are investigated using the linear method multinomial logistic regression as well as the nonlinear methods random forest and multi-layer perceptron networks. Results from nonlinear methods for dimensionality reduction do not detect any distinctly new patterns or clusters, compared to linear methodology. However, some results are promising to build upon in further methodology development. The best performing classification method shows results corresponding to wellknown effects for 70.6% of the compounds evaluated. Moreover, classifications of 11.8% of the compounds indicate potentially unknown effects, which are considered interesting and could be a springboard for further analysis and innovation. Therefore, this classification methodology can create insight and potentially high value.
- PostPlaytesting Match 3 Games with PPO(2023) Malec, Stanislaw; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Andersson, Adam; Haghir Chehreghani, MortezaThe training of proximal policy optimization agents with action masking on stochastic match-3 environments is explored in this thesis. A performant, feature-rich match-3 simulator is developed, and experiments demonstrate improved performance over a random policy on both seen and unseen levels. Furthermore, the best generalization performance is achieved when training is done by sampling levels from a subset of levels.
- PostUsing language models to improve a speech recognition based maritime emergency call detection system(2022) Johansson, Eric; Chalmers tekniska högskola / Institutionen för matematiska vetenskaper; Andersson, Adam; Röshammar, Kristoffer; Zechner, NiklasNovel applications of the transformers architechture as well as the availability of pre-trained models have drastically reduced the amount of data required to train successful speech-to-text (STT) models. By using the Connectionist Temporal Classification (CTC) algorithm, the process is further simplified as the training data does not have to be pre-segmented. This work aims to improve the performance of such a model developed to detect maritime VHF radio emergency calls by adding a language model to the CTC-decoding. We experiment with language models trained on several different text corpora and apply language models both in the decoding and on the resulting transcripts. The results indicate the importance of large amounts of domain-specific text. The results also show that a reduced Word Error Rate (WER) does not necessarily lead to an improvement in contextual comprehension. Finally, it is shown that relatively large improvements are given by fine-tuning various pre-trained STT-models on a curated dataset.