STATE OF THE AI - UPDATE
Greetings Summoners,
One of the most frequently asked questions during Summoner’s Fest was “When are we going to get an update about AI?” This remains a top priority for us, and we discussed covering it during the fest, but unfortunately, we decided that venue was better reserved for features that were closer to completion. That said, we do still want to update you on how it’s going and what’s next. But first…
A Short Recap on our Current AI
Our current AI knows only…
- What buttons it can press
- The distance between the two characters
- What animation state each of the characters is in
Our current AI does NOT…
- Have any memory of what happened earlier in the fight, or even in the previous frame
- Have any ability to plan for the future
- Have any idea what buffs, debuffs, masteries, nodes, etc. even are, let alone which are active
Our current AI is just a frame perfect button masher. There is no real intelligence behind why it chooses to do something; it’s just using weighted-random rolls to mash buttons until it finds a valid input, and it has no ability to learn. This also means that most “behaviors” players experience are an emergent property of its many weighted die rolls, rather than something that was intentionally designed. And if those behaviors weren’t intentionally designed, then they are functionally impossible to edit without wide-reaching consequences, which is why we started building something new.
If you want a more robust breakdown, check out our previous
AI dev diary.
AI 2.0 Progress
For better or worse, the current AI is the game, and players justifiably have lots of expectations around how it does (and should) behave. So for AI 2.0, our goal is to get as close to the current AI experience as possible initially, but built on a better, more interactive infrastructure. We know it won’t be exactly the same, but our goal is to get as close as possible, to make things better where it is different, and, most importantly, to have laid solid foundations for future improvements. So let’s talk about some of those foundations.
Perception System
Our first step was creating a robust Perception system that allowed us to analyze what actions the current AI could have taken on any given frame, which action it did take, and integrate that into our Fight Replay system. This allowed us to better quantify the current AI’s emergent behavior trends in data. Below is an example of that data with the measured weights removed. This gave us a starting point for tuning behaviors in our new system, and also gives us infinitely better insight into why the new system is making various choices than we’ve ever had with the current system.
Memory System
From the groundwork of the Perception system, we were able to build out a working Memory. AI 2.0 can now remember what happened earlier in the fight and use that to influence its decision making. As a quick clarification, this still doesn’t mean that the AI will be able to learn - we haven’t and can’t build Skynet - but if we ask it when it last threw a heavy, it will now know the answer. We won’t be using this system to reproduce current behaviors, but it is critical for future fixes, such as improving the experience of baiting out Special Attacks.
AI Logic Blocks
We’ve also created the ability to codify more complex behaviors, not just specific animations (Parrying, Intercepting, etc.) so that we can modify these individually in the future, without all the scary knock-ons. Our starting point for these blocks will usually be to reconstruct the existing emergent behaviors, for good or ill, but further changes will be much more predictable.
Proof of Concept - High Dex Bot
As an example of the kind of work we’re doing right now, and of why it doesn’t lend itself well to playable demos, here’s a terrifying example: This is what the AI looks like if we program it to try to avoid as many of your attacks as it can with Dexterity.
This is not something we ever intend to include in the game-wide AI, but it’s (perhaps unintuitively) the kind of thing that we have to build if we then want to prevent it.
Using another previous example, we can’t just tell the AI to “light intercept the Attacker less” unless it knows what that means. So we have to “teach it” to light intercept as perfectly as possible, codify that behavior in logic blocks, and then tell it to do that specific thing less (much less). We can’t just teach it halfway, or we’ll only be able to prevent it halfway. This work is absolutely critical for the kinds of changes both we and you want, but it weirdly means that much of our current work is building things we intend to block out, rather than to actively use.
What’s Next?
We know the current AI is frustrating; we feel it as hard as you do. We’ve been devoting resources to these updates for a very long time. The good news is that we are rapidly approaching broader internal playtesting on AI 2.0, meaning that we feel we’re close to the point where broader feedback will actually be helpful. We’re a little uncertain at this stage how much iteration beyond that it’s going to take to hit the “feels right” bar, so we can’t comfortably share a release timeline beyond “sometime 2026” at this juncture, but things are coming along, and this remains a high priority for the team as a whole.
However, for visibility, our intention is that once we feel it’s in a good enough state, we’ll be rolling it out for players to try in a piece of bespoke content (e.g. Side Quest), similar to how we initially tested things like Strikers and the Root mechanic.
We intend to build that content carefully to show off high-impact AI behaviors, and to let you see the new system in close proximity to the old. As we’ve said before, we don’t plan to just throw the switch game-wide without gathering as much broad feedback as we can. Maybe we get there and the consensus is to turn it on everywhere as soon as possible, which would be a great problem to have! But this isn’t something we intend to rush, because it’s something we intend to get right.
We look forward to sharing more details with you all as they become available!
