Brian Hourigan | University of Limerick (original) (raw)
Uploads
Papers by Brian Hourigan
Deep reinforcement learning has been extensively researched and applied to the realm of autonomou... more Deep reinforcement learning has been extensively researched and applied to the realm of autonomous vehicles, more specifically self-driving cars, over the last decade with varying degrees of success. It is well-established that there is no silver bullet to solving the conundrum of deploying safe and fully autonomous self-driving cars to real-life scenarios. This study seeks to evaluate three learning paradigms using a collection of minimalist autonomous driving scenarios thus exploring the ability of artificial intelligent agents to successfully learn challenging driving tasks. The three machine learning paradigms - online reinforcement learning, offline reinforcement learning, and imitation learning - are trained, evaluated, and generalized to negotiate the continuous and discrete action space-driving scenarios. By exploring and deploying a wide range of algorithms, techniques, and mechanisms, we show that a data-driven approach to learning from an expert agency rather than a direct interface with the environment shows an improvement in the overall performance of the autonomous vehicle. Lastly, we demonstrate the difficulty of common driving tasks in the context of autonomous vehicles by providing key metrics such as near-optimal rewards, collision metrics, and ability to generalize to new conditions to corroborate our analysis.
Deep reinforcement learning has been extensively researched and applied to the realm of autonomou... more Deep reinforcement learning has been extensively researched and applied to the realm of autonomous vehicles, more specifically self-driving cars, over the last decade with varying degrees of success. It is well-established that there is no silver bullet to solving the conundrum of deploying safe and fully autonomous self-driving cars to real-life scenarios. This study seeks to evaluate three learning paradigms using a collection of minimalist autonomous driving scenarios thus exploring the ability of artificial intelligent agents to successfully learn challenging driving tasks. The three machine learning paradigms - online reinforcement learning, offline reinforcement learning, and imitation learning - are trained, evaluated, and generalized to negotiate the continuous and discrete action space-driving scenarios. By exploring and deploying a wide range of algorithms, techniques, and mechanisms, we show that a data-driven approach to learning from an expert agency rather than a direct interface with the environment shows an improvement in the overall performance of the autonomous vehicle. Lastly, we demonstrate the difficulty of common driving tasks in the context of autonomous vehicles by providing key metrics such as near-optimal rewards, collision metrics, and ability to generalize to new conditions to corroborate our analysis.