3 Applications

There are many many different applications for ANNs. We will describe only two: one about playing a game and one about controlling a vehicle.

3.1 Encoding the inputs and outputs

Before applying an ANN, some attention needs to paid to how the input and outputs of the ANN are to be used. Note that both inputs and outputs are real numbers, and that each output is limited to the interval [0,1].

If the application requires the output to be a single integer, how are the outputs of the ANN to be used? One possiblity is to round a single output from the ANN to the nearest integer. However if the integer is itself a code for a set of unorded values, then perhaps it is better to have one output for each of these unordered values; the output of the ANN is then interpreted as the output with the largest value.

As another example, the outputs can be intepreted as estimated probabilities. If the probabilities must sum to one, then the outputs from the MLP will require normalising. These probabilities may be used as the basis for sampling from a set of disctinct objects (one per output).

These issues surrounding the input and output coding of the ANN are often key choices affecting the success of failure of its application.

3.2 The game of Nim

Nim is a simple game, trivial for a game theorist, but used here to illustrate how an ANN may be used to play a game.

Two players begin the game with 15 "beans". They take turns removing beans. On each turn a player may remove 1, 2 or 3 beans. He must remove at least one, and may not take more than 3. The player removing the last bean loses.

A player of Nim must be able to decide what to do when presented with 15, or 14, or 13, etc. beans down to two beans. This decision could be in the form of a simple table, and in fact the following table shows how to play the most perfect game possible:

A player following this strategy will always win if he has the first move. If the opponent has the first move, and if the opponent deviates from this table, then the player will also win. If both follow the table exactly then the player who goes first always wins. The x means that it doesn't matter; you will lose anyway. Actually this is true only if the opponent plays perfectly. If the opponent does not follow this table, then some values of x would be better than others, but this would depend upon the opponent's behaviour.

To use an ANN to play Nim, we will need to encode the input to represent the state of the game, and encode the output to represent the desired move. One possibility is to encode the input as a single neuron accepting an integer between 2 and 15 inclusive, and the output as a single neuron interpreted as one of (1,2,3) depending upon the closest of (0,0.5,1) respectively.

3.2 Robotic vehicle control

Another example is robotic vehicle control. Suppose there is some kind of autonomous vehicle, like a mars lander, that has sensors to measure its speed, direction, and distance from each of three radio beacons. These 5 measurements could be the inputs to an ANN. The outputs of the ANN could control the steering wheels and drive motors of the vehicle in order to accomplish some type of pre-defined objective, such as returning home to a recharging station.

Back to top. >>>

Continue with the tutorial. >>>