Previously, Artosis has made a great stream playing against some of the best Brood War bots using the SCHNAIL Client. If you haven’t watched that yet, here is the link:

The interview contains spoilers (duh). Originally I wanted to batch this with my analysis of the games, but that will be a separate post.

First of all, thank you for the stream! It was very exciting, and we learned a lot. Mostly what to fix 🙂 How was the whole experience? You seemed to enjoy it throughout, only making The Serious Face in the last match.

Hi! Yeah, it was actually quite fun. I was really surprised both by how the bots felt human, and also the extremely inhuman moves that they were able to make. It was really refreshing after more than 20 years of only playing terrible computers and humans.

What skill level did you expect from the bots at first? How did they measure up to your expectations?

I wasn’t exactly sure. I think in my mind, I expected them to be quite weak. I’ve played vs computers many times in the past, and they have always been laughable, so it was hard to imagine them being strong. That being said, when I saw that Cadenzie had lost maps to a bot, I was really impressed. She’s quite a good player, so that made me think that I should take the games quite seriously. I also had the opportunity to play against AlphaStar at Blizzcon, and it crushed me, so I knew losing was a possibility.

How did you find the SCHNAIL client itself? Is there anything you would add, change, or remove?

The client was great, very easy to use. It might be cool to have a way to put notes in or keep track of matches vs bots / replays, but honestly its quite nice.

Sneak peek from the next patch notes: “Science vessels can no longer cast hallucinate, no matter how much they want to”. Do you feel a strange void where there was something before?

Hahahaha, it was pretty funny when the Vessel had “Hallucinate” (which was just EMP with a different name).

Let’s talk a bit about the individual matches! Locutus sadly derped out (to be precise, it performed a “seal clubbing” tactic which is useful against weak bots. Last I checked you are neither weak, or a bot, so it predictably failed), but other than that, I think they gave us a good showcase. Krasi0 was playing well, but at some point it seemeed as if it couldn’t really decide what to go for. Do you think that match could’ve went any other way at some points? (To be honest, I don’t)

The way Locutus played reminded me of just a stronger BW computer player. It screwed up at my wall and couldn’t recover. That’s generally how games are vs PCs in StarCraft. Krasi0 was really funny to play against. It was comical having the vultures stay out of range constantly, even across the whole map. Each time I scanned, the Wraiths ran out of range too. That was CRAZY. I couldn’t believe how much stronger its Wraiths were than a human player’s.

I think Microwave and Marian Devecka gave us some good shows, they were just simply not on your skill level. Would you agree with that assesment, or add some other insights to it?

Hahaha, well  the mass Zergling play from Microwave was kind of funny. Once you identify what’s going on, it becomes pretty easy to defeat. Marian Devecka looked and felt the closest to playing a human. It went for Mutalisk harassment while expanding, micro’d and harassed my main, and even left at the right times or approximate right times to try and stop my Marine pushes. It was kind of creepy how human it was. There were some issues though. It made some additional Sunkens which looked like they were anti-Vulture Sunkens, but are useless and a waste against the strategy I used. Humans have very thought out scout timings which tell them if Vultures are a possibility, which I guess the bot just doesn’t have.

My impression is adias was a strong, but very predictable bot, and it also kinda employed the current meta – you were well prepared for it. Still, it is a strong bot, and not to be underestimated – would you agree on these points? While developing adias, the Samsung team actually consulted some Korean pros – do you think it showed?

Yeah, definitely. The bot seemed like it was quite strong, but had a very hard time when I put it into the purely tactical position of breaking a Tank contain. If I hadn’t set mine up first, that game could have gotten out of control pretty quickly.

And finally, the McRave match! It went on for 38 minutes, and it seemed you really needed to concentrate to finally prevail. Can you talk about how you felt during and after that?

The McRave match was a rollercoaster for me. Throughout any game, my feelings about whether I’m winning and how far ahead / far behind I am continually change. For a very long time I was certain of victory and felt far ahead. Then I kept scouting Nexuses that had full Probe saturation. Based on the army sizes I was killing, this seemed absolutely impossible. It honestly felt like it was cheating hahaha. I think around the time that I scouted the 3rd main base that it had, I thought I had a very good shot at losing. Luckily for me, the bot has no actual finishing moves, so I was able to grind it out. I think that game really shows the strength of the bot economy. I’ve played against and cast the best Protoss players in the world, and this bot had a way better economy than anything I’ve ever seen.

I would dare to say that McRave made a bad decision (double/triple expanding after almost destroying your natural) and that cost it the game – if it just kept up the pressure, it could have overwhelmed you. Would you agree? (Keep in mind i didn’t re-watch the replay, I’m guessing you did 🙂 )

Actually I felt like that’s why it did so well. The bots don’t have good killing moves, so I generally wasn’t afraid of dying outright. The mass expanding though seems to be its strong suit. That’s what I had lots of trouble with.

Were you prepared for the eventuality of losing the game before the cast? Did you prepare a sad, but ultimately awkward speech to your fellow players?

Hahaha, no speech. I wouldn’t have minded losing. I did a lot of casting and analysis for DeepMind and AlphaStar, so I’m totally at peace with the concept of AIs being better than me.

Not all bots have surrender logic – I think that’s not entirely a bad thing. It is annoying when you hunt for the last supply depot, but we saw Locutus recover in one of the Cadenzie games after she basically thought it was over. What do you think of this issue? What would be a good balance point?

I think it’s fine for the bots to stay, especially if they are learning from it. Honestly that’s a human flaw – leaving too early, or thinking its over and slacking off. I do both probably weekly.

What is the most dangerous trick/tactic the bots used, or you think they could use vs. humans?

The speed and precision of micro is probably the scariest thing. Cadenzie mentioned that full range perfect Dragoon micro is insane. I didn’t have to deal with those things as much, since Siege tanks are the longest range unit in the game. On my side, things like Wraiths getting out before the scan hits were huge. Other than that, the economy management was very strong. The bots can simply get more money than a human.

What aspects of the bot you think authors should focus on in order to become more competitive against humans?

I’m not exactly sure how the programming works, but bots seem to be a bit bull-headed. Like most people, their strength is their weakness. McRave got so much and kept doing large attacks, but it had no finess at clearing my drop at top right, and it didn’t have any sort of kill move like a big recall. The two Terran bots that I had seemed very strong mechanically, but they didn’t really know the tactics about dealing with Siege Tank containment, such as running around, big flanks, mass dropships, etc. These are the holes that humans can accel (I think he meant excel) at and beat the bots with.

All in all, how would you estimate the skill of your opponent? You can say an MMR, rank, amount of good boy points, or however you’d like.

I think that the bots were good. I honestly think that most of the bots that I played would destroy most human players. I’m not sure what MMR to say, because they all have points that can be exploited. I really do encourage anyone who plays StarCraft and is reading this to give it a try though, I bet you will lose most of your games 🙂

That brings up a point we’ve been thinking about, like, hard, maaan. How to actually evaluate the bots’ skill? They work differently, and on SCHNAIL, they don’t play against each other, just like the humans with different skills don’t play vs. each other there. Traditional MMR/ELO rating systems presume that everyone plays everyone. Do you have any thoughts on the matter?

It’s so hard. I think you’d have to put the bots on the actual ladder for a real understanding of their level. Like their macro and micro is S++, their problem solving is F, their build orders are A….how does all of this come together for an actual skill level? I think it will vary widely based on who they play and how that person or AI plays against them.

Do you think SCHNAIL could be used as an effective training tool and/or sparring partner for someone who is trying to get better at the game?

I think there’s definitely potential there. For instance, in the Terran vs Zerg against the Mutalisks… if that bot had a couple changes, it could be a great practice partner.

Will you personally use it as a training tool, or just to have fun? Maybe a warmup match?

Hahaha, probably not. It was really fun to play against them before, but their play is just a bit too off to really help me prepare at the moment. I will probably play against the best bots in some matches from time to time though, just for fun on stream 😀

What kind of bot would you like to see developed in the future?

I love the idea of using a bot as a practice partner for certain situations. It would be very cool to see some aggressive bots come out, like 2 gate Zealot rush, or a Dragoon rushing bot, and practice holding in those situations. If you can do it vs perfect micro, then holding it even vs progamers should be doable!

Thanks for answering my questions, and for the stream – I can say with confidence that the whole Brood War AI community is thankful for it! Any closing thoughts you’d like to add?

Thanks! It was really fun, thank you to Sonko and Antiga for talking to me a lot about this stuff, coming on The Pylon Show, and helping me to get it all going! Cheers to the bot creators, you all did a great job!