Robotics and machine learning

Types of learning

Rote learning

Learning by memorisation

Off by heart

Caching

Form of rote learning

Everything has to be stored

Impractical in a complex scenario

Learning from advice

Very broad and ill-defined concept

Difficult to transfer to machine learning

Learning from experience

Improves performance over time

'Evaluation Function' was assigned a weighting or ranking

Has led to the impressive success of genetic algorithms in some domains

Learning from examples (inductive learning)

System is provided with positive and negative examples

Uses these to make generalisations and form concepts

Winston's Arch

Aim was to see if a program could develop a representation of the concept of an "arch"

Neural nets (artificial neural systems)

Represented in the connection weights of a highly interconnected set of simple nodes

Follows three phases

Training

Validation

Application

Explanation-based learning

Uses a combination of a single example along with domain-related knowledge to explain the example

Requires a goal concept, and domain knowledge

Represented in a way which can be manipulated using deductive and inductive logic

Learning by discovery

Learning by analogy

"Blocks world"

Number of blocks which can be piled on top of each other

Uses predicates, like Prolog

Example: ontable(b).

Planning

Find search techniques

Exhaustive searching

Heuristic searching

Industrial robotics

Vision systems