Two Ways To Think About Intelligence

1. Knowledge Intensive: Intelligence is all about knowledge. We have to find out how to aquire, store, etc..
2. Data Intensive: There is a lot of ambiguity in the world - the world is its own representation. We have to find knowledge in the world.

Example: Robot Control

Comments

• This is the simplest network possible (there are several optimisations possible if you change the representation).
• Units of the network are in competition with each other.
• This network works for every linear seperable function.

Problems

1. Learning is very slow.
2. Requires a teacher.
3. Gets stuck in local minima (just like reactive agents).
4. There is only little retention.
5. Works only for linear seperable functions.

Ways To Get Out Of Local Minima

Genetic Algorithm

Use not only one network, but use several networks with different initial values and find out which one is close to the global minimum.

Critical assumptions

1. It is possible to define a fitness function.
2. It is possible to measure the fitness for every individual of the population.

Basic ideas

1. Include the fittest into the next generation.
2. Combine some of the fitter ones to make the next generation better.

Questions

• If you have a fitness function, why not just use the function to calculate the optimum ? - Example: Layout 25 things in a room. We know the optimum (all 25 things layed out) but we do not have the layout. Try several layouts and measure how good they are.

Problems

• Which encoding ?
• How to calculate the fitness function ?
• Efficience ?
• It is possible that you will not find the minimum, if the number of individuals is small and time is limited.

Simulated Annealing

Basic idea

• Introduce some noise into the system (= shake it from time to time).

Example: Reactive Control With Simulated Annealing

• Goals: Robot picks up remote control and avoids obstacles.
• Sensors: Laser (find remote control), Sonar (find obstacle).
• If robot stops (ex. remote control lies on obstacle), "kick it" and begin from another start point.

Questions

• Is Simulated Annealing a slower method than Genetic Alogorithm ? - Every method has lots of parameters, therefore it is very hard to compare them.
• How do we know when we have reached the global minimum ? (Is it possible that we will be "kicked out" of the global minimum ?) - The noise function depends on the time. "Kick more gently as time goes by."
• Why do we need all that stuff ? Teacher based neural networks simulate the way humans learn. - We do not know if this is the way humans learn. There are also differences between teachers. If you want to add 8 and 3, a teacher could tell if you did it wrong (8+3=5 -- failure) or he could tell you the right result (8+3=5 -- failure, 11).

Multilayered Neural Networks

Questions

• Why do we need multiple layers ? - In CS we use modularity, data encapsulation and abstraction to handle difficult problems. We do exactly the same. If the problem is too complex: Decompose it !
• Why do we not just add tons of layers ? How do we know when to stop ? - If we decompose too much, the connections are getting complexer than the problem.
• Sounds nice, but what do the hidden layers mean ? - We do not always know what the layers mean (see also Tic-Tac-Toe).

Example: Tic-Tac-Toe Network

CS problems if programming Tic-Tac-Toe

• Rows
• Columns
• Game Over
• Empty Spaces
• Singlet
• Doublet
• ...

Meaning Of The Hidden Layer

This network is quaranteed to play Tic-Tac-Toe perfectly (after training). Every unit of the layer represents a concept (row, column, ...), but you can remove a unit and it will still be able to play Tic-Tac-Toe perfectly. If you remove a unit it is not possible to specify the meaning of the other units. This phenomen called "distributed representation of concepts"

Questions

• Why do we not use more layers (ie. put the groups of five pieces in eight layers) ? - Layers depend on each other. In this model the groups do not depend on each other.

ALWINN (Automatic Steering Of An Automobile)

Basic ideas

Camera (with low resolution 32x30) in front of the car. A driver teaches the network while driving (ie. network tries to predict actions and receives feedback from driver). Constant speed is assumed (cruise control).

The network uses sigmoidal instead of binary functions as input and output (steering wheel & pixel of camera).

Homework

Think about how this works, because the network has only little retention (ie. it might forget about driving straight while training for driving a curve).