Simple Models of Learning based on using 'raw' information

This chapter is a first pass at exploring models of learning based strictly on the input for visual perception (pixellated 2D images, with colour depth) and the output for action, in the form of specifications for muscle tension for each muscle.

Probably the most common concept of learning focuses on personal experiences and on repeating 'successful' action sequences more frequently than 'unsuccessful' action sequences. A similar type of learning is to avoid action sequences that lead to pain.

• The learning is based on discriminating between successful and unsucessful action sequences in prior experiences kept in memory
• The model assumes that there are relevant experiences
• The model does not address how the first relevant action sequences might have occurred
• The model looks at the basic information processing functionality required for such learning
• Goal setting, comparison: e.g. matching a present sensation of hunger against a previous experience of hunger that is followed by satiating hunger
• A second goal is to recognize circumstances and action sequences that led to feelings of pain.
• selection: finding an action sequence with a similar starting context that appeared to be most successful in leading to a successful outcome
• Alternatively, to avoid action sequences that led to pain.
• The model assumes the remembering and use of 'raw' information

Adapting behaviour to be more successful: utilizing perception to modify action

• The learning above looks at whole action sequences. Presumably we can use a similar approach to find better subsequences within a larger sequence. It makes sense to modify action by taking an extra step, or by turning to the right instead of the left, etc.
• This adaptation is still based on 'raw' visual experiences and on 'raw' action sequences