Explaining imitation learning: potentials in preprocessing, language for explanations and frame-wise importance
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Nowadays, beyond machine learning are foreseen to change the world, penetrating into every aspect of our lives. Imitation learning, a machine learning paradigm that learns from demonstration, demonstrates superb performance in field like autonomous driving and robotic manipulation. While more and more complex models integrate into imitation learning, the opacity also increase. In this case, this thesis aims to investigate the explainability in field of imitation learning, which is rarely researched by the research community. We propose three frameworks that explore various ways that \tmpnt{implements} eXplainable Artificial Intelligence (XAI) into imitation learning, while maintain the overarching goal of enhancing the explainability and policy performance at the same time.
This thesis presents three key contributions. First, the GenIL framework enhances imitation learning through genetic operations for data augmentation, enabling reinforcement learning to extrapolate rewards from suboptimal demonstrations with reduced exploration costs. By improving extrapolation accuracy and consistency, GenIL optimizes the use of limited data and integrates XAI to clarify genetic operations for stakeholders. Second, the Ex-PERACT framework introduces interpretable robotic manipulation using a hierarchical transformer model that integrates natural language and visual inputs. It learns abstract skill codes for multitask execution while providing intuitive explanations, improving human-robot interaction and system effectiveness. Finally, the R2RISE framework enhances imitation learning explainability by identifying critical frames in sequential inputs, offering model-agnostic, frame-wise explanations that improve transparency and trustworthiness for stakeholders.
The aforementioned research demonstrates comprehensive methodologies for addressing challenges at the intersection of XAI and imitation learning. The studies exhibit significant advancements, from uncovering the potential of XAI in imitation learning, to proposing model-specific explanations under language conditions, culminating in the development of a model-agnostic framework applicable to a wide range of imitation learning models. Initially, we focused on enhancing the efficiency of imitation learning with limited demonstrations, highlighting the potential of explainability in improving model performance. Subsequently, we bridged the gap between human language instructions and agent predictions using a hierarchical transformer-based model. Finally, we investigated frame-wise explainability across various models. Collectively, these studies provide effective solutions to the opacity issues in imitation learning, thereby contributing to the overarching goal of optimizing explainability in imitation learning decision-making processes.
Please use this identifier to cite or link to this item: