Starcraft AI blog

As reported, ISAMind is a fork of Locutus, and the only important difference is that ISAMind can predict the opponent’s opening plan using a trained neural network instead of the rule-based method that Locutus inherited from Steamhammer (and modified). The author of Locutus has conveniently made a branch for ISAMind, so that we can compare and verify the differences.

The neural network is a standard feedforward network trained by backpropagation (we can see this in the network data in the file ISAMind.xml). The computation is implemented using OpenCV, a computer vision library that includes some general-purpose machine learning tools. I’ve looked at OpenCV in the past, and I think it is a good choice.

The input to the network is the frame count and the counts of any early game units seen, plus any opening plan recognized already so the network can decide whether to stick with it, plus the high-level features of the rule-based recognizer for proxy, worker rush, factory tech, and number of known enemy bases. With those high-level features, the network doesn’t have much thinking to do. The output is one of 10 possible opening plans, and if none is recognized it falls back on the rule-based recognizer again. In the best case, this neural network can’t provide much of a boost.

The key question is: How was the network trained? I looked all around and found no sign of an explanation. It could be trained to reproduce the values returned by the original rule-based method, but what would be the point? It could be trained to recognize the plan that leads to the highest win rate, but that would be expensive and might learn to deliberately misrecognize plans, so that’s less likely. It could have been trained on data from games with opponents that were coded to play specific plans, but that would risk lack of variety in the training data. It could have been trained on hand-labeled data, if they took the time to label that much data. My best guess is that they wrote a replay analysis tool that labels the training data “here is what I saw” with the plan, “here is what I should recognize,” that would have been chosen if the scout saw everything. The scout is early and normally does see everything that is in the enemy base (not proxied in some hidden location or built later outside the main), so if my guess is right, the trained network should normally be accurate.

I ran ISAMind locally to see what the predictions looked like. My impression is that the predictions were generally accurate, but sometimes sluggish. Steamhammer recognizes a zealot rush if it sees 2 gateways and no gas. ISAMind doesn’t recognize a zealot rush until it sees the zealots. Whether the sluggishness is harmful depends on whether the bot needs to react quickly. ISAMind recognizes a fast rush at the first sign, and that is the most important enemy plan to recognize immediately. Most input information has to come from the scout probe, and the scout probe circles the base without ever looking outside, so the network rarely sees enemy expansions unless it passes them on the way in; the input data is not complete enough to recognize all the possible plans reliably, because scouting is not thorough enough. (Locutus has a commit “Scout the enemy natural” only later, on 19 September.)

Overall, ISAMind seems like a cute little job that doesn’t try to accomplish much. It is like a class project or an early experiment to start getting tools and ideas into place.