Goal Recognition (GR) techniques aim to infer the intentions of an autonomous agent according to the observed actions of that agent. We introduce the evidence-based GR framework, regarded as the primary contribution of this thesis, which is designed to address goal recognition problems in both static and dynamic environments. Furthermore, we leverage the proposed framework to develop GR techniques aimed at addressing challenges in the transhumeral prosthesis scenario.
The evidence-based GR technique learns knowledge models using process discovery techniques. It then constructs conformance diagnostics between the learned models and a new observed action sequence executed by an agent. These diagnostics are then used to formulate a probability distribution over a range of possible goals of the agent, where the probabilities indicate the likelihood of the observed action sequence for achieving each possible goal candidate. The evaluation results confirm that the evidence-based GR approach grounded in process discovery and conformance checking techniques studied in process mining achieves comparable recognition accuracy to other state-of-the-art GR approaches and reacts faster. Notably, distinguishing itself from other GR approaches that rely on handcrafted domain knowledge, the evidence-based GR approach can automatically learn knowledge, making it applicable across diverse real-world scenarios. Additionally, the learned knowledge models are explainable, compared to models used in deep learning-based GR approaches.
This thesis emphasizes the challenge of GR in non-stationary environments, where the GR system is required to continuously solve GR tasks over time, during which the underlying environment may change. An adaptive GR framework is proposed as an extension of the evidence-based GR framework for static environments. This adaptive framework can detect changes in the behaviors of the observed agents and adapt the learned knowledge models accordingly. An evaluation of three specific solutions to the adaptive GR problem that are implemented as different control strategies of a GR system is presented and discussed. The evaluation is based on a collection of adaptive GR problem instances generated by the tool we designed and implemented, Goal Recognition Amidst Changing Environments (GRACE). The results demonstrate a trade-off between the GR performance over time and the effort invested in adaptations of the learned knowledge models, showing that few well-planned adaptations can lead to a consistently high GR performance.
To verify the usefulness of the proposed GR techniques in real-world scenarios, their applicability in a powered transhumeral prosthetic scenario is assessed, where it is required to detect a person’s intended movements based on electromyography and kinematic signals collected from their body. A powered transhumeral prosthesis aims to restore missing anatomical segments below the shoulder, including the hand, and is designed to assist patients with disabilities. It analyzes continuous, real-valued data from sensors to recognize patient target poses, or goals, and proactively move the artificial limb. Our GR techniques were evaluated using offline datasets and online human-in-the-loop experiments, comparing the results with state-of-the-art techniques such as linear discriminant analysis (LDA)-based and neural network-based approaches. The results demonstrate that the proposed GR techniques grounded in process mining achieve superior performance.