Learning research

Learning is the process of adapting behaviour in response to events in the environment. Without learning, a system is limited by the ability of its designer to foresee all situations that might occur. The aim of learning research program is to massively increase the range of operation and robustness of autonomous systems.

Research into Learning has three main elements:

1. Recognition

Recognition of features and situations in complex dynamic environments is a pervasive problem for intelligent autonomous systems. It requires the application of both perceptual and domain knowledge. Learning such knowledge greatly increases the robustness of the system.

Recognition research focuses on:

  • Supervised and unsupervised learning methods for clustering in feature space and the use of domain knowledge to guide feature construction.
  • The use of relational learning methods, especially inductive logic programming (pioneered by us), and Bayesian networks in building and maintaining representations of complex scenes, and
  • Research in recognition will fundamentally advance the ability of agents to understand a scene or operating environment.

2. Decision making

In dynamic environments, an agent must be highly reactive and able to make decisions rapidly. It is desirable to have a large library of behaviours that can be quickly matched to specific situations. The research question is how such libraries should be built. We have broad expertise in automated decision-making ranging from control-theoretic to symbolic methods.

Two approaches are investigated:

  • Symbolic learning using high-level languages that allow hybrid systems that combine reactive behaviours with deliberation through “any time” algorithms, and
  • Hierarchical reinforcement learning algorithms that decompose the learning problem into more manageable subtasks.

3. Multi-agent Learning

Teams of agents have the potential for accomplishing tasks that are beyond the capabilities of a single agent. An excellent and demanding example of multi-agent cooperation is in robot soccer.

Research in multi-agent learning focuses on:

  • Designing structures for learning in a hierarchy of goals.
  • Multi-agent, cooperative, reinforcement learning, and
  • Representations of belief-desire-intentions (BDI) of agents, auction algorithms and other mechanisms for learning joint utility from agent interactions.