Suction caissons have been used extensively for anchoring and supporting the offshore installations like oil platforms and wind turbines. These foundations are normally subjected to complex combinations of the vertical, horizontal and moment loads (i.e. V, H, M) from the self-weight, wind, wave and currents. In the past decades, extensive studies have been conducted to investigate the combined V-H-M loading behaviour of suction caissons in clay. However, most existing studies are focused on the ultimate bearing capacity, while the deflection response is more critical in foundation design for recent infrastructures like offshore wind turbines. Due to the complex load conditions, predicting the three-dimensional (3D) deflection response of the foundation is still challenging. Machine learning (ML) appears on the research horizon due to its excellent capacity of solving nonlinear problems with desired speed and accuracy. However, conventional machine learning approaches were limited in their capacity to analyze raw natural data without artificial interventions. Meanwhile, the deep learning technique (DL), as a branch of machine learning, allows a machine to be fed with raw data, automatically extract the features, and discover intricate structures in high-dimensional data. The deep learning technique has been used in many fields like language translation, auto-pilot and image recognition. And Deep neural networks, including deep learning algorithms and architectures, are gradually being developed. In light of these backgrounds, this study proposed to develop a deep learning based surrogate model to predict the 3D deflection response of suction caissons under combined V-H-M loading. The advanced three-dimensional nonlinear finite element (FE) simulations under complex V-H-M loading paths were performed on suction caissons of different geometric configurations and in clay soils with different stiffness and strength properties. The 3D FE simulation data was then used to train the deep learning based design model. Three popular neural network structures, i.e., Feed forward Neural Network (FNN), Convolution Neural Network (CNN), Recurrent Neural Network (RNN) have been employed to develop the hybrid surrogate design model. In this study, two different training strategies were proposed for this geotechnical problem. In the first category, the 3D load-deflection behaviour of suction caisson is idealized as a point-to-point mapping problem, i.e. mapping between the deflections (i.e. displacement and rotation) with loads (i..e force and moment). This task was achieved by Fully-Connected Neural Network model (FC-NN) based on FNN, One Dimension Convolution Neural Network model (1D-CNN) based on CNN and Long Short Term Memory model (LSTM) based on RNN. In the second training strategy, the load-deflection response was idealized as a time series process, a line-to-line mapping problem, mapping between the past loading paths (i.e. 10 groups of forces and moments) with future loading paths (i.e. 90 groups of forces and moments). Besides the three neural network models mentioned before, another two complex and advanced models, LSTM Model combined with convolution neural network (1D-CNN+LSTM) and Temporal Convolutional Network model (TCN), are also applied for temporal prediction. The performance and training efficiency of these models were also systematically evaluated by interpolation and extrapolation experiments. Basically, all the models can well capture the 3D deflection response of the foundation with significantly high accuracy (i.e., root mean squared error is smaller than 0.05 and coefficient of determination is near 1.000) than the traditional design approach (such as macro-elements model), and with greater efficiency than the 3D FE simulations. Among all the models, the TCN model has the highest prediction accuracy and robustness. However, the FC-NN model has the simplest model structure and highest computational efficient in learning the non-linear relationship between deflection response and V-H-M load. Besides capturing the relationship between input and output, the deep learning model can also assist to identify the intrinsic failure mechanism. By observing the fluctuation of generalisation ability, the evolution of the failure mechanism of suction caisson with embedment depth was revealed.

Model-free reinforcement learning has proved to be successful in many tasks such as robotic manipulator, video games, and even stock trading. However, as the dynamics of the environment is unmodelled, it is fundamentally difficult to ensure the learned policy to be absolutely reliable and its performance is guaranteed. In this thesis, we borrow the concept of stability and Lyapunov analysis in control theory to design a policy with stability guarantee and assure the guaranteed behaviors of the agent. A novel sample-based approach is proposed for analyzing the stability of a learning control system, and on the basis of the theoretical result, we establish a practical model-free learning framework with provable stability, safety and performance guarantees.
% Specifically, a novel locally constrained method is proposed to solve the safety constrained problems with lower conservatism. In our solution, a Lyapunov function is searched automatically to guarantee the closed-loop system stability, which also guides the simultaneous learning (covering both the policy and value-based learning methods). Our approach is evaluated on a series of discrete and continuous control benchmarks and largely outperforms the state-of-the-art results concerning unconstrained and constrained problems. It is also shown that the algorithm has the ability of recovery to equilibrium under perturbation using the policy with stability guarantee. (Anonymous code is available to reproduce the experimental esults\footnote{\url{https://github.com/RLControlTheoreticGuarantee/Guarantee_Learning_Control}}.) Since sometimes the constraint is hard to define, we introduce a novel method to learn a constraint by representing the bad cases or situations as a distribution, and the constraint is the Wasserstein distance between the distribution.

Neural networks have achieved great success in many difficult learning tasks like image classification, speech recognition and natural language processing. However, neural architectures are hard to design, which requires lots of knowledge and time of human experts. Therefore, there has been a growing interest in automating the process of designing neural architectures. Though these searched architectures have achieved competitive performance on various tasks, the efficiency of NAS still needs to be improved. Moreover, current neural architecture search approach disregards the dependency between a node and its predecessors and successors.
This thesis builds upon BayesNAS which employs the classic Bayesian learning method to search for CNN architectures, and extends it to the problem of neural architecture search for recurrent architectures. Hierarchical sparse priors are used to model the architecture parameters to alleviate the dependency issue. Since the update of posterior variance is based on Laplace approximation, an efficient method to compute the Hessian of recurrent layer is proposed. We can find candidated architectures after training the over-parameterized network for only one epoch. Our experiments on Penn Treebank and WikiText-2 show that competitive architectures can be found in 0.3 GPU days using a single GPU for language modeling task. We find that our algorithm is more efficient than state-of-the-art.

The internal model is an important piece of the control system of an autonomous driving vehicle. In order for the model to deliver accurate predictions, a valid model structure and well chosen parameters are needed. Model parameters can be highly fluctuating or complex to predict, especially when looking into tyre ground surface interaction models. Instead of predicting parameter values beforehand, they could be estimated and updated in real-time. Fluctuation or incorrectness can be adjusted while driving. However, this uncertainty in parameter value must be accounted for when applying control. Solving this problem by regarding the uncertainty in parameters of the internal vehicle model as a POMDP has been researched in this paper. The research question being: is it worthwhile to use the POMDP approach for online parameter estimation of autonomous passenger vehicles? To answer this multiple sub-questions have been composed. We start off looking into: what is the most suitable vehicle model? Different vehicle models and tyre models were compared. Literature showed the bicycle model in combination with the linearized tyre model to be most suitable for autonomous passenger vehicles. The next question is: What is the most promising algorithm? Using literature, suitable algorithms for solving this POMDP have been found and compared. From three compelling algorithms, the one best fitting the autonomous driving criteria was chosen. Knowing the model and the algorithm for the simulation the next question became: Does the algorithm perform on a vehicle model? To answer this question, the simulation has been implemented in MATLAB and performance has been tested. The results showed significant increase in parameter estimation performance. Within 2 timesteps the estimate had converged correctly. The next question is: Does the algorithm perform within realistic bounds? To answer this question, the same simulation as before has been used, but now with saturation on the steering input. This showed parameter estimation performance increase compared to the original trajectory, but not as overwhelming as without saturation. The next question is: Does the algorithm suffer from high noise? To answer this question, the same simulation has been used, but now with different levels of noise. The results showed parameter estimation performance significantly affected by increasing noise. The final sub-question is: Does the algorithm suit increasing model complexity? To answer this question, the amount of parameters have been increased in the simulation and there has been looked into the large matrices that accompany the algorithm. Results showed that increasing the complexity has a significant effect on the size of the simulation and algorithm matrices. In conclusion, from all of these experiments arose some very interesting results. This produced a useful insight into the strengths and weaknesses of the POMDP algorithm performing on a passenger vehicle, answering the research question. This also led to various recommendations for future research. Interesting would be altering the belief filter to enhance performance on bounded steering input and increased sensor noise.