- Toggle the add obstacle button to draw obstacles. Remove an obstacle by clicking it with the middle mouse button.
- Click the play button to start the learning process
- hide graphics disables the rendering of the dots, improving the calculating speed of the algorithm
- Run best shows you the best performing dot up untill this point
- best of each generation loops through all the generations up untill this point, showing you the best performing dot of each generation (press right arrow key to move to next generation)

In the bottom center of the screen you can see the brain of the best performing dot. It is represented by the input-layer (the leftmost column containing the bubbles numberd 1 - 9) and the outputlayer (bubbles 10 and 11). The brain will evolve based on the situations the dot encouters.

The inputlayer (1-9) represents the dot's senses, of which it has nice. It can detect which direction it is currently facing and it has vision in 8 directions: up, down, left, right and diagonal.
The outputlayer (10 and 11) represents the actions the dot can undertake. In this implementation the dot can do two things: steer (10) and accelerate (11).
The dot's job is now to relate the inputs on the actions it can undertake. For example: ff the dot detects an obstacle to it's right while it is traveling north, it should learn not to steer to the right since this would result in a collision and death. So the dot has to learn what is the 'best' action. But how does it decide which of the actions are 'good'? That is where the fitness function comes in.
The fitness, or 'how well does this dot perform?' is pretty important. It is used to make a selection out of the 1000 dots that have spawned select the best performing dots for 'breeding'. Just like in Darwin's theory of evolution well-performing dots have a higher change of survival. See the NEAT paper for more details on how the algorithm works under the hood.
In this example the performance of each dot is dependent: if the dot does not see the goal (there is a straight line between the dot and the goal without any obstacles inbetween) then the fitness increases the more distance the dot travels. If the dot does see the goal then it's fitness is determined by the distance to the goal (smaller is better). This ensures that dots that have not seen the goal start searching and dots that have seen the goal have a greater change of being selected for breeding.

For questions and/or suggestions to this implementation you can leave a comment on this post