• Machine learning for building intelligent tutoring systems
The Client : Javna agencija za raziskovalno dejavnost RS
Project type: Research projects ARRS
Project duration: 2011 - 2014
  • Description

It is generally accepted that one-to-one tutoring is by far more effective than in-class teaching, but too costly in most situations. To carry out one-to-one tutoring at reasonable cost, CAI (Computer Assisted Instruction) is an alternative. To overcome CAIs rigid behaviour, ITS (Intelligent Tutoring Systems) are generally considered as the most promising option. However, the costs of ITS development are high because it requires major involvement of a domain expert to realise intelligent student-system interaction. The purpose of this project is to develop methods for automating the process of developing ITS and thus considerably decrease the high costs of building ITS, by enabling more economical, semi-automatic development of ITS.

In particular, the main goals of the project is to develop methods to support automated conceptualisation of learning domains, which can be viewed as a key component in the construction of ITS systems. In the project, several experimental case studies in selected learning domains will be carried out from two types of domains: symbolic problem solving, like physics, medical reasoning and chess, and motor and/or control skill domains like running, walking in parients, or tennis.

The role of domain conceptualisation is as follows. In complex domains, the connection between the basic domain theory (axioms, laws, formulas, rules of the game, etc.) and problem solutions is usually rather complex and hard for a human to execute. Therefore, there is typically a need for an intermediate theory, conceptualised domain theory, that serves as a bridge between the basic declarative domain theory and procedural knowledge for concrete problem solving. This can be viewed in terms of a derivation chain as follows. Basic domain theory (axioms, etc.) logically entails a conceptualised domain theory, which in turn entails problem solutions. Logically, the basic domain theory also entails problems solutions, but at much higher problem-solving effort. The basic theory is typically non-operational for a human (requires excessive computation, or it may be too complex to memorise), whereas the conceptualised theory is human-assimilable. These relations can be illustrated as follows:

original theory ------------------------------------------> problem solution

original theory ----> conceptualised theory ---> problem solution

A conceptualised domain is a problem solving tool for a human, and therefore it should be simple and compact, so that it can be understood, memorised, and executed in problem solving by the student.

The planned conceptualisation methods will be based on some recent methods and paradigms of AI which will be adapted and further elaborated for the purpose of this project. These techniques and paradigms include: ABML (Argument Based Machine Learning) , QR (Qualitative Reasoning and modelling) , Q2 learning (Qualitatively faithful Quantitative learning), EBG and EBL (Explanation Based Generalisation and Learning), ILP (Inductive Logic Programming), and specific techniques of behavioural cloning (capturing human skill from observed humans behaviours).

The developed conceptualisation methods will be experimentally applied to intelligent tutoring systems for selected domains as follows:

(1) symbolic problem-solving domains: physics, medical diagnostic reasoning (diagnosis of tremors), and chess;

(2) motor skill domains: running, tennis, walking in patients (multiple sclerosis)

The AI part of the project will be carried out by the Artificial Intelligence Laboratory of Univ. of Ljubljana, and LMG laboratory of the Ljubljana Clinical Centre will collaborate in the neurological medical applications. These medical applications are expected to be directly adopted in the medical practice.

/* */