You can train him to drown himself by moving him to a water square, and repeatedly rewarding him.

Watch the blue connections grow on the graph of his memory to show he is associating the water with pleasure.

After a while you can stop, and watch barry eat the water, and eventually drown himself.

Please do not worry, Barry isnt completely stupid, he will "relearn" that drowning is painful, and move back to the red berrys. ( though you can use this same technique on the berrys too).

Simulation controls:

Take control of Barry:

Train Barry:

Pressing G will cycle the graphs, one of them shows Barry's brain, the neural network.

The network has inputs;

Red Berry, White Berry, and water. Which describe the square Barry is on.

And the output is the single neuron in the thrid layer.

This is Barry's predicted pleasure for eating the current square.

The red lines denote negative effects on activation, and the blue denote positive effects on the following neurons activation. The width of the line shows the strength.

So you can watch how Barry learns by looking at this graph.

The first layer of 'Neurons' (Red B, White B and Water) are the input neurons and are activated depending on what square he is on.

The one neuron in the third layer is the output neuron. If this is activated, this will make Barry eat.

It represents the pleasure(or pain) Barry associates with that square. If he predicts it is pleasurable, he will eat, if not he will look at the surrounding squares and move to the one he deems most pleasurable.

How I want to take this further :

Neural networks are adaptable, they can also be trained to do complex things like work out cross products. I want to create more intelligent AI by creating modules of Neural Networks that do different tasks, which come together to produce AI for games that adapts to repeated strategies used by the player, but will also adapt to changes made to the game as the game develops in production.