what kind of game is your bot actually playing
Every game rewards something specific. Most bots are built for the wrong one.
-
-
making your bot worse on purpose
I was stuck trying to make the bot feel more human. Then talking to this game director changed my approach
-
When to commit your Vikings against Tempests
Tempest are a pain and Tyr shows a simple rule for your bot to engage high range targets
-
The map feature your bot is ignoring
Why Hello Reader, I was watching a replay where a single marine held off two roaches for way longer than it should have. If you can believe it was because the marine was standing on a Xel’Naga healing shrine and getting 10 HP back every tick. Most bots don’t even know healing shrines exist. Dogtato walked through this on the latest PostBots session. The shrines are sitting right there on two maps in the current pool. Torches has one that allows air units. Persephone has one for ground units…
-
I couldn’t tell if my bot was actually improving
I stopped guessing whether my bot was improving. The answer was never in the game I kept watching.
-
Your bot lost. Do you know why?
Why hello Reader, I watched my bot lose a game last week. Stared at the score screen. Knew something went wrong around the three-minute mark but couldn’t tell you what. Sound familiar? Telemetry replaces the guesswork. When your bot loses, terminal print outs tell you what happened during 1 game, telemetry tells you what happened over many. I built a dashboard that layers my bot’s decision data on top of replay analysis. Battle sim confidence over time, army value curves, worker counts,…
-
why I stopped reading the terminal
Why hello Reader, I spent way too long trying to get a Zealot to stand guard for my Zerg rush defense. He had one job: move into the correct location when an enemy appeared. But spending hours squinting at terminal logs, trying to juggle X and Y coordinates in my head just to figure out where he actually stood, was driving me nuts. I was in terminal hell. It’s a miserable way to build. When you try to debug pathing by reading a wall of text, you’re taking something that happens in space and…
-
they think it’s just a game bot
Why hello Reader, “No one takes bot making seriously.” Was the response I got on Discord from a member telling me about his struggles applying a new role, after I asked why his SC2 bot never makes it into his LinkedIn posts. I sat on that for a while. Then, with the World Cup coming up, Boston Dynamics drops this awesome breakdown of their Atlas robot playing football. They ask a singular question: “Can football teach a robot to move?” They’re basically asking if a game can actually train a…
-
How to train your bot’s micro from scratch
Why Hello Reader, Training micro with reinforcement learning sounds like something only Alphastar could do. But what if you could start with just one unit and two actions? SharkNice showed how to easily train unit micro from scratch using Sharky RL Matrix Training (or Stable Baseline 3 for python) , starting with a single Stalker kiting two Roaches and scaling up from there. But the key take away though : scope your action space to preset functions your bot already knows. Instead of giving RL…
-
Bots keep ignoring the past
Why hello Reader, I like to be right. Cause who wants to be wrong? 🤣. But that leads to missed opportunities because I talk myself out of taking action, waiting for certainty. And the exact same thing happens to my bot. Remember when I talked to you about setting my bot up to predict cheese? Well, it works a good amount of times. But when it’s unsure, it delays making a decision until it’s too late. We already know that writing endless IF statements for every edge case is a trap. We also know…
