I’m so happy to see this! Thank you so much for the AI progress. I’ve been an avid VS AI player since it gives me the ability to play the game without feeling any serious pressure.
The improvements have come a really long way as I remember steamrolling AI in the blink of an eye. Now, there’s been a few tight spots based on team composition that make it really interesting at the Elite level.
That said, I would like to see the AI push a little more outside of Objectives. It’s very obvious that they wait for minions when in their lanes otherwise they remain at the gates. It’s essentially how we find out who’s in our lane is just getting right at the point of tower targeting and see what AI is sitting there.
They also have predictable pathing towards Objectives and rarely, if ever, engage a human player when they’re on the way to one. If you cross paths with AI towards a focal point, they just keep moving for their goal instead of trying to gank the human player.
i only play versus a.i.
and i must say that current elite ai is always deciding not to kill.
i mean most of the time when im sure im dead for some reason they decide to fallback.
and sometimes they stuck at move command at certain areas of maps.
it was fine at beta and 2.0.
but you guys did something and it became like this.
Why… would you sabotage yourself to increase the chance of your team winning? I mean, sure, you could come back just before the game ends, but if you’re not even going to play the game portion of the flipping game, why play?
Sounds like you’re looking too much into this. A.I. is still an intuitive portion of the game, not just with leavers and disconnects, but just learning the game. The A.I. in heroes has been abysmal; so bad that you really can’t learn anything from them, and the gap between A.I. and PVP is HUGE.
Sounds like a good approach.
I know it’s probably not your job, but is it possible to communicate this bit of info in the game itself? It’s such a significant change, even the players that don’t read patch notes must know it.
At least put a link in the launch screen to the patch notes (maybe even a blog post?) explaining the changes and the reasoning behind them.
I’d love to see something like this eventually, but as I understand it there would several problems in the way.
The first and most glaring would be that HotS just isn’t a big enough IP to do this with quite yet. Google wanted a complex but also popular game, and SC2 was perfect for that. HotS is nowhere near as popular nor complex as SC2 is. OpenAI chose DOTA2 because that game is extremely popular, and there are a ton more minor variables to account for (due to the existence of items).
A lot resources would need to be pumped into this project, and HotS is so far down on the totem pole of Acti/Blizz’s priorities that a massive collaboration like AlphaStar just won’t get the funding it’d need.
Another is fine-tuning the skill level of the AI. AlphaStar can’t really pull it’s punches, so you can’t tell it to tone it down for any given opponent; it will always do it’s absolutely best to win. As a result, AlphaStar’s iterations are either elite or trash, with almost no in-between. That means we can have Masters and pro-level bots, but even AlphaStar isn’t advanced enough yet for anything below that that could give even a Bronzie a run for their money. Trying to do the same thing in HotS will likely be just as difficult.
There’s also the length of time going into this project. AlphaStar took about a year to produce any real results. I suspect HotS would get going faster due to its relative simplicity, but we’d still probably be looking at a half year or longer just to get to this “AlphaStorm” bot up to the level of AlphaStar.
You’re missing the entire point here, I was weighing between making AI somewhat good that it would be able to beat actual human team or leaving it in a completely useless state that players will still be able to steamroll against it in less than 5 minutes. Realistically you can only achieve the first by programming bots never to miss skillshots and optimally utilise their cooldowns, which only makes it frustrating to play against.
Bots will never miss out on XP soak by intentionally waiting in the bushes for 30 seconds only to ambush you, they will never be able to create a diversion on one end with a boss or merc camp only to attack another side to waste your time running across map. There is no situation that playing against bots will prepare a new player for. Bots will stay in constant movement between two actions which only makes landing skillshots against them annoying because of how erratic their patterns become.
Also for the people glorifying Blizzard man over his appearance, ask him who made the mess that the current AI system has become and why they simply didn’t revert the changes which were almost instantly disliked by the entire community. As it stands the AI in year 2016 was drastically better than what we have now. Last three times they were boasting about improvements, it turned out to be even worse, like no one got the chance to jump in three games to playtest it before releasing to the public.
If I sound a bit rude, I won’t apologise for the lack of blind faith, because I am very bitter about how WOW 8.3 and WC3:R turned out. Instead of making poor attempts at positive PR, try releasing the patch on PTR as soon as possible so we don’t experience the fiasco of Deathwing patch all over again. You know what I mean, the patch that was finalized for Blizzcon but it still took 5 weeks to materialise on the PTR with no changes at all. Two weeks after leaving the live servers in broken state, the anomaly effect was changed almost entirely because of the negative and “unpredictable” effect it had on gameplay.
Hello Mr. Gu, i understand a lot of work has gone into developing the A.I. decision tree but i cannot help but wonder, why use a primitive system like this that will constantly require work to be up to date?
There is no way to non-arbitrarily define the value of a pushed lane and compare it to the value of a team fight for the objective for example.
So, my question is i guess, why not simply use deep learning and avoid a thousand hours of work in the future?
When someone goes afk in a gold game, the level of understanding of the AI that has to replace that guy can non-arbitrarily be determined since we can save a copy of the AI each time it’s elo goes up, so we would have an answer for every case and there would be no doubt that we have answered correctly, since we can measure the elo of the AI directly.
I am of course not saying that making a huge decision tree is not a good idea in a vacuum; only that when compared to the deep learning, it seems to be much more work and much less good.
For example with the current decision tree, the way i understand it, you will have to teach the AI how to use every single talent in the game basically and that will require a lot of time and knowledge from your part.
I just find deep learning to be the simpler more elegant solution.
Can I just say – as someone who has played 3000+ games versus the AI over the course of the past 5 years and who loves that mode dearly – thank you.
The evolution of the AI since the inception of HotS is quite remarkable. Every time it has been improved I have gotten a better game experience playing against it. And even though it has its flaws and weaknesses, it is actually a great joy to play against.
I can’t wait for the next improvements to happen, and I shall look forward to seeing what AI behavior changes it leads to.
P.S: Thank you for making the AI an absolute beast at playing Maiev. There’s no greater joy in life than being brutally murdered by its laser-precise Fan of Knives over and over again.
I wouldn’t mind it if the difficulties for the vs A.I. Was more tuned to something like “Bronze, Silver, Gold, Diamond-Master” So if I pick Elite, it would almost be like playing against higher tier players vs picking Easy being like playing against Bronze.
And I do hope that what ever changes are happening to the AFK A.i. replacement isn’t going to make them WORSE then they already are. I’m a little concerned because it sounds like you are planning to make the AI replacements dumber…? If I read that right.
I don’t know about Versus A.I, but talking about DC/AFK I aways appreciated the AIs (except when then circle around, but that’s kinda rare… for me at least). I mean, a Half over a Zero does a pretty good job and they usually take the blame instead of the DC’d (which are the real problem in these cases but also not for this post).
I think something the A.I lacks is some sort of a sense of safety based on the hero it is controling which should change a bit how it’s Goal works: Healers should try to stay close behind their team prior to going to objectives or laning. Heroes who doesn’t have a decent waveclear shouldn’t be waveclearing when not urgent. the A.I should know to avoid enemy territory (which change a lot with laning, vision, health and visible enemy heroes) specially when they’re alone and, with that being said, they should know how to regroup before advancing to an objective or such.
Of course I don’t know how these translates to game design and I don’t know if these suggestion would even be viable workwise.
If you think something I said is dumb or unnecessary or if you think it could be adjusted somehow I’m very open to discussion ‘-’