Don’t Expect AI to Play Like a Human

The debate surrounding AI and StarCraft shows that we need to change the way we discuss and evaluate artificial intelligence — and stop comparing its gameplay with our own.

By Ben Dickson

DeepMind’s recent exploits in developing artificial intelligence that can defeat world-class players at StarCraft II caused a lot of stir. While DeepMind called it a major breakthrough, others argued it was cheating, unfair and superhuman.

But what the entire debate indicates is that perhaps we need to change the context in which we discuss and evaluate the capabilities of AI — and stop comparing its gameplay with our own.

AlphaStar, DeepMind’s StarCraft-playing bot, uses deep learning, a popular field of AI in which programmers develop the behavior of their AI models by giving them an insane number of examples. AlphaStar first trained on a large database of human game data released by Blizzard, playing millions of games against itself to learn and master the rules of StarCraft. It was then pitted against humans, sweeping DeepMind’s own amateur players before going on against world champs.

When AlphaStar beat TLO and MaNa, two of the world’s best players, there was reason to believe the artificial intelligence industry had passed a milestone. In a blog post, DeepMind called AlphaStar “a step forward in our mission to create intelligent systems that will one day help us unlock novel solutions to some of the world’s most important and fundamental scientific problems.”

But then came the criticisms.

An Unfair Advantage

Critics claim that AlphaStar has several characteristics that make it an unfair opponent against humans.

First, AlphaStar is blazing fast. DeepMind engineers say they handicapped AlphaStar to prevent it from performing more actions than a human could accomplish. But human players do a lot of spam clicking, or impulsive actions that have no value or thinking behind them.

For instance, when players want to order their units to move to a location or attack an enemy, they often click repeatedly on the same location or on a trajectory toward the destination, because it gives a false feeling that clicking will speed up that action. In reality, the units execute only the most recent command and will ignore previous ones. In contrast, AlphaStar’s every move is precise.

Critics argue that the mismatch lets AlphaStar micromanage the game in a superhuman manner. For example, in a large battle where many units are involved, AlphaStar can give individual commands to each of its units with speed and precision that would be impossible for its human opponents. In an analysis of AlphaStar’s performance, ArsTechnica’s Timothy B. Lee described a few scenarios in which AlphaStar’s speed and precision would give it an unfair advantage.

Other analysts have pointed out that AlphaStar receives more information than human players. The version of the bot that beat MaNa and TLO had access to the entire map, as opposed to seeing a monitor’s worth of battlefield space like a human player. But it was still limited by “fog of war,” which means it couldn’t extract information from the areas where it did not have active units.

Yet others criticized AlphaStar’s limits: It could play only as Protoss, one of the three races in StarCraft, and in only one of the many maps of the game. Given a new race and map, AlphaStar would probably lose against amateur human opponents because, from the AI’s perspective, it would be like playing a totally different game.

What Is Fair Play?

DeepMind still hasn’t released technical details, but some suspect that instead of having to process raw pixels as humans do, AlphaStar might have had access to raw game data through APIs (application programming interfaces).

Ars’ Timothy B. Lee comes to this conclusion: “The ultimate way to level the playing field would be to make AlphaStar use the exact same user interface as human players.” This means that, like a human player staring at a computer monitor, the AI would have access only to the game’s graphics and have to simulate keystrokes, mouse clicks, and scrolls instead of interacting with the game through API calls.

This would be a fair point if we expected AI to replicate the human brain and senses exactly. But deep learning and neural networks, which are still the cutting edge of AI, have distinct limits that prevent them from reproducing some of the most basic human functions.

Deep learning is narrow AI, which means it’s very good at performing specific tasks such as labeling images or recognizing speech, but it’s awful at generalizing tasks or transferring its knowledge to other domains. The more you broaden the problem domain, the more limited the AI’s capability becomes and the more training it needs. That’s why AlphaStar won’t be able to play another RTS game, such as Warcraft 3 or Company of Heroes.

It also took AlphaStar 200 years’ worth of games to master Protoss on a single level. It would probably take just as much to learn to play Terran or Zerg, the other two races of StarCraft. In contrast, a human player could quickly port over the knowledge they gained from one game to a new one.

We are still decades away (at least) from general AI, the type that can match the cognitive capabilities of humans. Some scientists believe we will never succeed in reproducing the human brain.

But narrow AI is very good at processing large amounts of information at very fast rates. That’s why AlphaStar can handle the entire map of StarCraft at the same time. The designers of StarCraft could have modified the game to provide players with a full view of the game map, but that would probably confuse the players rather than help them. Humans can also be given access to raw game data, but that too would be of no help.

Humans are slow at processing data but have common sense and abstract-thinking capabilities that enable them to plan and make decisions without complete information. That’s why they prefer to have a limited view of the map and to focus on a single part of the battlefield; at the same time, they have a sense of what’s going on in other parts of the game and can develop a general game plan.

Is AlphaStar Cheating?

Given the differences between AI and the human brain, it’s fair to say that the critics were right in their assessment: DeepMind rigged the competition in favor of AlphaStar by limiting it to a single map and a single race. But the debate about AlphaStar can bring us to some very important conclusions.

First, the main point of the game shouldn’t be to check whether AI can click and scroll like a human. Instead, we should focus on how AI performs in a game that provides imperfect information and requires real-time decision making. In this regard, AlphaStar did a pretty good job.

Second, StarCraft might not be the best venue to test the strategizing and planning capabilities of AI. As one analyst pointed out, “StarCraft II is a game that can be broken by mechanical perfection.” This means AI can compensate for its poor strategy skills with its superhuman speed and surgical precision.

Finally, AI and human intelligence are so different that it would probably be impossible to create a level playing field between the two. The smallest changes to the rules would quickly tilt the game in favor of one side or the other to a degree that would make the competition unfair.

We should look for environments and settings where we can unleash and test AI to its full potential instead of slowing it down with artificial human limitations. What could humans and AI achieve when they’re cooperating instead of competing?

Originally published at www.pcmag.com.