Dimensions of Machine Learning in Diagnosis

1. What can trigger learning?
   • Failure; Success; Differences between expected and real values
   • Need to improve abilities

2. What are the elements supporting learning?
   • Critiques, praise, estimates, evaluations and advice (internal or external)
   • Sequences of diagnosis decisions
   • Diagnosis histories, e.g traces of information flow, knowledge exchange, negotiation
   • Analyses of failures and conflicting elements (goals, decisions)
   • Feedback after completing the diagnosis task

3. What might be learned?
   • Constraints relating parameters or other elements of the diagnosis
   • Dependencies between diagnosis parameters
   • Support in favor of or against a decision
   • Diagnosis rules, methods and plans
   • Analogical associations
   • Preferences
   • Preconditions and postconditions for rules, actions and tasks
   • Consequences of diagnosis decisions
   • Types of failures and conflicts
   • Heuristics for failure recovery and conflict resolution
   • Successful diagnoses and diagnostic processes

4. Availability of knowledge for learning
   • Direct communication (with the user or another diagnosis system)
   • Indirect communication (e.g., between diagnosis systems via a blackboard)
   • Record of the state of the diagnosis
   • Repositories of diagnosis and interaction histories

5. Methods of learning
   • Explanation-based learning
   • Induction
   • Knowledge compilation
   • Case-based and analogical learning
   • Reinforcement learning
   • Genetic algorithms
   • Neural networks

6. Local vs. Global Learning
   • Learning by a diagnosis program
   • Learning by a group of diagnosis programs

7. Consequences of learning
   • Diagnosis improvement
   • Improvement of the diagnostic process