Towards Trustworthy Decision-Making in Human-Machine Symbiosis
Type of DegreePhD Dissertation
Computer Science and Software Engineering
MetadataShow full item record
As artificial intelligence (AI) evolves, it becomes an integral part of our daily lives. To augment our effectiveness, human-machine symbiosis enables both humans and AI systems to offer different yet complementary capabilities. However, one of the significant concerns in human-machine symbiosis is the lack of human trust due to the potential ramifications, risks, or even dangers caused by AI. The critical question here is no longer whether AI will have an impact but by whom, how, where, and when this positive or negative impact will be felt. Trust is a prerequisite for humans to develop, deploy and use AI. Without AI being demonstrably worthy of trust, its uptake by humans might be hindered, hence undermining the realization of AI’s vast economic and social benefits. This dissertation centers on building human trust in AI approaches to sequential decision problems, i.e., trustworthy decision-making. Specifically, there are three significant issues in current approaches. (i.) The first issue regards robustness where the brittleness in the planning indicates its inherent weaknesses. This identifies the potential risk that the AI system is unreliable and may lead to a blind trust that an AI system stays prone to errors even with high performances. To address the issue, I developed a framework to equip planning with the ability to learn so that the representation used for planning can be improved through the learned experience. Experimental results on benchmark domains demonstrate that the proposed approach can adapt to the domain uncertainties and changes and improve reliability. (ii.) The second issue regards interpretability where the learning behavior of deep reinforcement learning based on black-box neural networks is nontransparent and hard to explain and understand. This is identified as one of the main barriers to building human trust in the outcomes produced by the AI system. I developed a framework to address the issue by leveraging task decomposition and causal reasoning. Therefore, the task-level system behaviors can be interpreted in terms of causality -- causal relations among different sub-tasks. Experimental results on the challenging domain with high-dimensional sensory inputs empirically validate the interpretability of sub-tasks, along with improved data efficiency compared with state-of-the-art approaches. (iii.) The third issue regards adaptive autonomy where the concern is to what degree of autonomy should be granted to an AI system. Furthermore, keeping humans in a supervisory role is key to striking a balance between machine-led and human-led decision-making. Therefore, I developed a human-machine collaborative decision-making framework to empower the machine agent to make decisions, with humans maintaining oversights. In addition, the openness supported by this paradigm, i.e., the willingness to give and receive ideas, can also increase human trust. Experiments with human evaluative feedback in different scenarios also demonstrate the effectiveness of the proposed approach.