Reinforcement learning (RL) has grown tremendously over one and a half decades and is increasingly emerging in many real-life applications. However, the application of RL is still limited due to its low training efficiencies and surplus training cost. The sampling and computation complexity normally depends on the size of the state space and splitting the state space can distribute computation and accelerate learning. State abstraction as a form of data-centric method shrinks the state space and reduces learning time, however, it is challenged by the fact that abstraction throws away information and might result in a sub-optimal solution. In this thesis, we propose the hierarchical clustering-based state grouping (HCSG) method to split the ground state space into clusters and train multiple agents for each cluster without changing the dimension of the state space. This approach allows us to distribute computation and improves training efficiency without losing the overall performance, and was also shown to outperform baseline and other state-of-art data-centric methods.

Semi-autonomous driving innovations aim to bridge the gap to fully autonomous driving by co-operating with human drivers to lead to optimal choices on who should drive in different scenarios by offering different automation levels. However, in the present day, known semi-autonomous driving solutions do not generalise to every complex case of driver and AI interaction. This limitation prompted research in attempting to solve the problem using artificial intelligence and machine learning techniques. This paper focuses on providing a reinforcement learning approach to solve one specific decision-making scenario of the driver initiating a shift of control to a different automation level. The decision problem was formulated as a Markov Decision Process, and the problem was solved both by a baseline handcrafted decision tree and a learned reinforcement learning policy using the DQN algorithm. The two policies were compared based on safety, comfort and efficiency metrics in a simulated driving environment. The results were indicative that a reinforcement learning policy generally ensured safety \& comfort and has shown increased efficiency over the baseline policy, however, it faced efficiency & comfort issues in outlier cases.

Transitioning to use automated vehicles is a gradual process. Until full automation capabilities are developed there is a need to mediate which driving entity - human or autonomous driving system (ADS) - should be in control depending on the circumstances. This research aims at investigating the switching between manual and automated driving in semi-autonomous vehicles when the ADS becomes unfit to drive. To this end, a simple environment simulation was created and an MDP model was formulated that accounts for sensor failures and leaving the operational design domain (ODD). Deep Q-Network (DQN), a deep reinforcement learning (RL) algorithm was trained and evaluated against a hand-curated decision-tree-based standard. The DQN-based policy did not reach the performance of the baseline algorithm. The conclusion is drawn that using DQN to handle this multi-objective decision problem using an intuition-based reward function cannot learn an optimal policy.

Over the last two decades, autonomous driving has progressed from science fiction to a real possibility and rapidly developing. However, autonomous driving technology has significant weaknesses and is not safe in unexpected conditions. As a result, automobile manufacturers insist that the driver remains in the driver's seat even while the vehicle is in autonomous mode. Semi-autonomous driving helps in this situation. Semi-autonomous vehicles require minimum human intervention and cooperate with human drivers. It provides multiple levels of automation to the driver to give the optimal decision about who should be in charge in a particular scenario. However, it has limitations and does not work perfectly in all complex scenarios. This paper focuses on this limitation and provides a reinforcement learning strategy to solve an existing complex scenario. In this specific scenario, Mediator initiates a shift of control to a different automation level when uncomfortable driving situations are detected. Markov Decision Process strategy used to formulate the decision problem, and the reinforcement learning strategy compared with a decision-tree based baseline strategy for the evaluation. The outcome was collected using driver safety and comfort metrics. The outcome supports the hypothesis, demonstrating that a learned reinforcement learning strategy can be used to solve complex decision-making scenarios in semi-autonomous driving.

Automated driving is a rapidly growing technology nowadays. Semi-automated driving is a subpart of automated driving which has multiple driving modes where both driver and automated module can take control. But full safety and comfort guarantees cannot still be given to the drivers. In this project, research has been done to ensure driver safety and comfort for the driver on the decision logic module of a semi-automated vehicle. Research focuses on just one specific use-case which is driver fatigue. The main goal is to ensure driver safety and comfort in the case where the user is sleepy during manual driving. Markov Decision Process is used to create a model to successfully represent this case. Evaluation has been done using already existing Reinforcement Learning algorithms and comparing their performances with the decision-tree based baseline policy. In conclusion, a successful Markov Decision Process model is created. While evaluating, some models performed better than expected and some are worse. In the end, a successful model is proposed, it can still be developed further to provide more safety and comfort for the driver.