Explainability in human-artificial intelligence collaboration
Επεξηγησιμότητα στη συνεργασία ανθρώπου-τεχνητής νοημοσύνης
View/
Keywords
Discrete Soft Actor-Critic (SAC) ; eXplainable Artificial Intelligence (XAI) ; Explainability ; KernelSHAP ; SHapley Additive exPlanation (SHAP) ; Deep Reinforcement Learning (DRL) ; Human Robot Collaboration (HRC) ; Gazebo Simulation ; Robot Operating System (ROS) ; Human–AI interaction ; UR3 cobot ; Co-learningAbstract
Recent advancements in Artificial Intelligence (AI) allow the development of AI methods that enable human-AI and human-robot collaboration. Here, collaboration is defined as the operation of humans and robots in a shared workspace to achieve common goals. Deep Reinforcement Learning (DRL) methods have been used recently for human-robot collaboration tasks allowing also co-learning of the task at hand. However, such methods, despite their ability to generate desirable behaviors, they cannot explain agents’ behavior, i.e. why an agent decided to select a certain action over all other possible actions. The lack of explainability in AI methods has many implications, including reduced trust in human–AI interaction, as well as issues of transparency and accountability. In the present thesis, we use SHAP values as a method to explain the agent’s behavior by analyzing how the features of the state space affect its selected actions. In our case, an agent is trained through a Soft Actor-Critic (SAC) DRL algorithm to learn how to collaborate with a human in order to collaboratively move a UR3 cobot towards a common spatial goal in real-time, in a simulation environment. Furthermore, we use KernelSHAP to compute feature importance within the agent’s state and leverage these values to generate explanations based on data collected from an expert human–agent team. Finally, we check whether the SHAP values align with human intuition.
Postgraduate Studies Programme
Τεχνητή Νοημοσύνη - Artificial IntelligenceDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνCorporate Department
National Centre for Scientific Research "Demokritos"Number of pages
99Language
EnglishCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα