Ask a question about League or Riot, and we’ll try to answer it. Answers go live every other Thursday at 1:30 pm (PT)
Looking back on AlphaGo’s victory over Lee Sedol a year ago, it’s clear that game AI has come a long way since we first implemented our existing bot AI. We are curious to see how far we can take artificial learning networks in League.
We knew we wanted the AI to be able to select champs, make strategic decisions during games, and have a unique learning pool for every champion. We started looking into this with a convolutional belief hierarchy which compresses a spike-and-slab pointer network. It can be difficult to calculate the intelligence of an AI, but, by our approximations, done through a sparse dictionary regression of the QRS-factors in the underlying transport structure, we believe that these bots could soon approach 200 IQ.
The bots are designed to learn from players in real-time. This is why we’re pulling test games from PvP queues instead of co-op vs. AI games; we want the bots to have the most realistic PvP experience possible. We’ve created an exclusive icon to reward players who participate in and finish one of these games before April 2nd, 2017, at 11:59 PM. (Icons can take a few weeks to appear in your account.)
We’ve seen some really interesting behaviors so far, with Advanced Bots sometimes emoting after kills, spamming taunts, or attempting to trick opponents. They’ve even developed their own distinct personalities and playing styles, likely due to phenome mutations within the Markov expressions of individuals due to backpropagation of the combinatorics resulting from the directed acyclic graphs.
But how exactly is it all done? I’m no expert, but the basic concept is rooted in a supervised learning process which accepts input sequences consisting of real (but Fourier reverse-transformed) block-chains generated through synthetic non-guttering. The demodularization of outlying Craighton values from left-side variance stabilization yielded significant gains in both qualified and non-qualified planar transepts.
The team initially struggled with the gamma-type convexity inherent in the network fabric of the central neuralities inhibiting the magnification of the sinusoidal p-nodes, but were successful in mitigating the concern with on-device filter drivers. Looking forward, the team plans to migrate all existing Chebyshev equations used for calculating real-time Gringel coefficients to the much more efficient wave-mechanical approach to infinite Z-wells.
Hopefully that sums it up. Thank you for participating in our Advanced Bot testing, which launches today.