We proposed Error reduction merging (ERM), a new learning method that automatically discovers similarities in the structure of the agent’s environment. ERM identifies different types of objects solely from the data measured and merges the observations of objects that behave in the same or similar way in order to speed up the agent’s learning. We performed a series of experiments in worlds of increasing complexity. The robot had to learn a qualitative model predicting the change in the robot’s distance to an object. The results indicate that ERM was capable of discovering structural similarities in the data which indeed made the learning faster, clearly superior to conventional learning. This observed trend occurred with various machine learning algorithms used inside the ERM method.
COBISS.SI-ID: 9731668
The argument-based machine learning paradigm was adapted for usage in educational setting. In this paper, we demonstrated that by automatically selecting relevant examples and counter examples to be explained by the student, the ABML knowledge refinement loop provides a valuable interactive teaching tool.
COBISS.SI-ID: 10703956
We presented a novel approach to program synthesis that can be used as a basis for automatic hint generation in programming tutors. Instead of using a state-space representation of the problem-solving process, our method finds a set of textual edits commonly used by students on program code. Given an incorrect program it then synthesizes new programs by applying sequences of edits until a solution is found. The edit sequence can be used to provide hints with varying levels of detail.
COBISS.SI-ID: 10805076
Objective: The paper describes the use of expert's knowledge in practice and the efficiency of a recently developed technique called argument-based machinelearning (ABML) in the knowledge elicitation process. We are developinga neurological decision support system to help the neurologists differentiate between three types of tremors: Parkinsonian, essential, and mixed tremor (comorbidity). The system is intended to act as a second opinion for the neurologists, and most importantly to help them reduce the number of patients in the "gray area" that require a very costly further examination (DaTSCAN). We strive to elicit comprehensible and medically meaningful knowledge in such a way that it does not come at the cost of diagnostic accuracy. Materials and methods: To alleviate the difficult problem of knowledge elicitation from data and domain experts, we used ABML. ABML guidesthe expert to explain critical special cases which cannot be handled automatically by machine learning. This very efficiently reduces the expert'sworkload, and combines expert's knowledge with learning data. 122 patients were enrolled into the study. Results: The classification accuracy ofthe final model was 91%. Equally important, the initial and the final modelswere also evaluated for their comprehensibility by the neurologists. All13 rules of the final model were deemed as appropriate to be able to support its decisions with good explanations. Conclusion: The paper demonstrates ABML's advantage in combining machine learning and expert knowledge. The accuracy of the system is very high with respect to the currentstate-of-the-art in clinical practice, and the system's knowledge base is assessed to be very consistent from a medical point of view. This opens upthe possibility to use the system also as a teaching tool.
COBISS.SI-ID: 30199257
Conceptualizing procedural knowledge is one of the most challenging tasks of building systems for intelligent tutoring. We presented a novel algorithm for semi-automated conceptualization of procedural knowledge based on goal-oriented rules in symbolic domains. We applied the algorithm to the challenging KBNK chess endgame, and carried out a pilot experiment to evaluate whether the obtained concepts (instructions) could serve as a teaching tool. Somewhat surprisingly, even the beginner level chess players were able to quickly grasp the concepts, and learn to deliver checkmate. A separate, subjective evaluation of the instructions by experienced chess trainers was also positive.
COBISS.SI-ID: 9326164