I’m going to describe what we’re actually building behind the scenes. This will act as a teaser, whilst also providing more technical information for those interested. I’ll also provide a little more detail on how we’ll be building out the project, for those interested in the business side.
Our primary goal is to build tools that help competitive players improve. That is essentially the iTero slogan. Tools. Improve. Tools to improve. We started with the drafting coach, which is now become a market leader in that somewhat niche aspect of League of Legends.
Since the start of this year our focus has shifted onto what we call an “AI Coach”. Maybe, to be more specific I would call this an “AI Macro Coach”. It will study your recent games and identify patterns in your gameplay. Specifically, it is looking for the patterns that, in aggregate, most contribute to you losing . In other words, what do you regularly do that decreases your chance of winning a game.
To do this, the first thing we need is a lot of data. It can’t just be the typical “Kills/Deaths/Assists”, although that is where we start. To do this, I watch a lot of my own games. I ask myself “What did I do wrong?” and then try and translate that into data.
For example, I watch a game where I lose because my lane opponent snowballed and carried the game. It all stemmed from me dying twice to two separate jungle ganks. I create a new metric: Death to Jungle Gank. I can now measure all the games that I die this way and determine whether it’s a regular issue across 100s of games, or was just an unlucky game. I also adjust this by the Champion’s I play, since some are more likely to die than others. Then we rank this against other players, since it’s all relative.
Then, it’s onto training the AI. We feed the model millions of games and include all the data we’ve gathered. These range from how many dragons a team took to how often an individual players died under a turret, and many more. The model is rewarded when it correctly predicts who will win the game based on the information it’s given, and so it trains itself to maximize its understanding of how they all contribute to the final outcome.
For some statistics this is easier than others to work out. The team that takes the most inhibitors has a much higher chance of winning the game, somewhat from the pressure these cause but also because of the fact you literally can’t win a game without taking at least one.
However, what happens when we look at something like the number of vision wards placed? Low vision is, of course, bad. However, no one thinks only spending Gold on vision is a good idea, either. There’s a trade-off and AI is particularly good at finding the optimal values. Even better, it can find these optimal values given other information. Should you buy more or less wards when you are behind? When you are ahead and your team is behind? How about when the game is even? Each state requires a new optimal that would be impossible for the human mind to accurately calculate.
The reason I am so confident that is is a genuine game changer is because I’ve been using my own account to test it. I found out more about my own playstyle than from all the replays I’ve watched (and it’s a lot). The model called me out and said “You are really, really bad at trading at the start of the game”. It dawned on me, this AI-inspired epiphany, and there was no denying it. The statistics were clear and unemotional - I was taking almost 20% more damage than I dealt within the first 7 minutes in almost all games. I was in the bottom-10th percentile of all players on this one statistic. I want to climb, and now I have a solid gameplan. Rather than typing in “LoL Guide” into YouTube I can type “LoL Guide Early Laning Phase Trading”. I can also track this statistic overtime and in every game. I have something outside of simply Winning or Losing to focus on. As long as I’m trading better in the first 7-minutes, I’m getting better. This is the future of coaching.
I’ve said it countless times before and I’ll say it again for the record: this can’t replace a human coach. If you can afford to regularly pay $30-$50 a session then getting a real-life human to watch your games and give precise feedback is always more valuable. However, most of us can’t afford this. An AI Coach can be made available for free to everyone. If I was to explain to a new player how to become pro I’d give them this:
- Play the game, lots.
- Watch generic guides to get a base understanding of the game.
- Use general statistics to make informed decisions on Champions and Items.
- Use an AI Coach to determine improvement areas, work on them (through both video content and general practice).
- Hire a real-life coach to provide hyper specific advice, preferably a coach who’s comfortable using the data from the AI to help them be more refined.
How close is it to ready?
Hopefully from this, you’re be at least a little excited by the idea. So, how long before it’s available for all to use?
We have almost finished an end-to-end deployment. That means, there’s an app we’ve built that loads your data, runs it in the AI and displays it for you to view in a way that’s understandable.
Our next step is two-fold:
- Use it ourselves, a lot. We’ll do whatever we can to stress test it, whilst also finding areas that we want to improve on.
- Spend more time adding more data and improving the AI model.
These two can happen simultaneously and at a guess, I would say it is 3-weeks of work. Then, we roll it out to a small set of specially selected friends and advisors for further feedback. Again, it’s a case of listening to their feedback and making changes or fixing bugs as needed. Call it another 3 weeks.
Then, we go into an Open Beta through the iTero Discord (link). We’ll slowly phase this up, looking at a total of 100 or so users for another 4-6 weeks. Once there’s no more bugs to fix or obvious changes to make, we’ll roll it out live to the general public. If you’re tallying up, you’ll see it takes us to around September - which is when we hope to go live.
From a business perspective, we’ll stay relatively quiet at first. We’ll have 25,000 of the existing iTero users to test it out first. I imagine there’s at least 8-weeks of feedback and bugs to work on before we start bragging about it. After that, well - let’s not get ahead of ourselves just yet.