Search basics
- Problem states: initial, goal
- Expander function
- Goal tester, estimator functions
- Some kinds of search
Parameter search
- Hill climbing, gradient descent (ascent)
- Error (goodness) as a differentiable function of each parameter
- Local and global minima (maxima)
- The shape of the error surface, e.g., how spiky it is
Neural networks, evolution, and search
- Neural networks
- What are the problem states?
- If this is gradient ascent, what do we use for the error/goodness measure?
- For supervised learning, some function of the difference between the target and actual
output for each output unit
- But what about the hidden units? Back-propagation provides an answer.
- Biological evolution
- What are the problem states?
- Each individual's set of traits
- Parallel: each individual specifies a different problem state
- Relatively uninformed: the closest we have to an estimator function is death
