I've found some tricks to get it to learn really well (by gen 20 they all finish really fast) 1. Using a relu, return value => value >= 0 ? value : 0.0; 2. Bump

I've found some tricks to get it to learn really well (by gen 20 they all finish really fast) 1. Using a relu, return value => value >= 0 ? value : 0.0; 2. Bumping it up to 30 cars (if you use too many, it slows down, and you dont get as good of results because each iteration takes ages) 3. Changing the lap condition to c.lap < 5 and the time to 30 s. This lets each race go longer and widens the discrepancy between the best and worst racers - which improves selection