Machine Learning

• Model of Learning Agents
  • Machine learning (ML) is used to populate the KB of an intelligent agent. This knowledge is used to guide search.
  • Distinguish between intelligence and learning
  • Many different learning algorithms and schemes
    • Decision trees
    • Neural networks
    • Inductive logic programming
    • Genetic algorithms
    • Self modifying algorithms
  • Machine learning is often done in advance on training data, and the resultant model deployed for use in an AI agent. It is also possible to (continue to) learn during the real activity of the AI agent.
  • The learning is expected to generalize from the cases presented (as training data) to new unseen cases.

• Forms of learning
  • Unsupervised learning
    • Learn similar situations.
    • Can be used to learn groups of states, and use to choose transformations which work well for the group.
  • Supervised learning
    • Learn input-to-output mappings from preexisting cases.
    • Can be used to learn outcomes of transformations, with the state and transformation as input, and the result (acquired by the sensors) as the output. Once the mappings are learned they can be used to predict outcomes. This is useful for non-deterministic problems.
  • Reinforcement learning
    • Learn the quality of an action by evaluation from critic.
    • Can be used to learn which transformations to use.

• Road Map (for me mostly)
  • Decision trees - learning classification (supervised).
  • Neural networks
    • Counter-propagation networks, Kohonen layer - learning classification (unsupervised)
    • Counter-propagation networks, Grossberg layer - learning mappng (supervised)
    • Back-propagation networks - learning functions (supervised)
  • Genetic algorithms - finding/optimizing solutions (supervised)

Exam Style Questions