Neural Network Co-op vs AI
When I saw the april fools article, at first, I didn't realize it was an april fools joke. I seriously thought they were going to create a neural network AI.
Who thinks this would be cool? Imagine a new co-op vs AI mode (maybe it replaces intermediate). Every time a game is played, data is collected, and then at some interval (days? a week? a patch release?), the next generation of bots would be released. In a year, we could have bots that even challenger players couldn't defeat!
Typically, neural networks are limited because they require tons of trials to actually get anywhere. So, for anything moderately complex, you need millions of simulations, which would require a whole server to run them. ...which Riot already has setup and running. If they limited the mode to level 30 accounts (to better teach the bots), it could offer better IP rewards than intermediate bots, so people would have incentive to play it if they were going to play bots anyway. (Which I do quite often, to de-stress after a ranked loss.)
For those of you who don't know what a neural network is, here's a summary, along with what it might look like in league. Basically, you have an AI with a bunch of inputs. These inputs are connected to outputs in various random ways. You start out by completely randomizing the ways they are connected, generating some number of test AIs. Then, once all of them have run the program, you take the ones that performed the best, add some small random differences (called mutations) to the ways the inputs are connected to the outputs, and you start over again. Each batch of AIs is called a generation. So in league, the inputs would be where every (visible) champion, minion, monster, and ability is. The outputs would be movement, auto attacks, abilities, items, and summoner spells. A generation could be a day's worth of AI games, and the ones that performed the best would continue on to the next day, along with a bunch of AI that are similar but with one or two random mutations.
