Giant open source data and release of research tool "StarCraft" have become the key to general intelligence

On August 7, Facebook's AI team released the largest "StarCraft: Battle of the Mother Nest" game dataset in history. Immediately afterwards on August 10, DeepMind and Blizzard jointly released the game dataset of "StarCraft 2" and the AI training platform SC2LE.

There were many big events in the AI world last week, but only one thing attracted the interest of two AI giants at the same time, and that was "StarCraft."

On August 7, Facebook's AI team released the largest "StarCraft: Battle of the Mother Nest" game dataset in history. Immediately afterwards on August 10, DeepMind and Blizzard jointly released the game dataset of "StarCraft 2" and the AI training platform SC2LE.

Why have AI giants such as DeepMind, Facebook, and Ali become fascinated by the research on "StarCraft"? Why didn't they create an artificial intelligence to defeat human players, but instead embarked on the path of open source data and publishing research tools?

In addition, I believe everyone who cares about AI has noticed that many media and experts believe that the research on "StarCraft" can guide AI towards the path of general intelligence. But most explanations don't know why. What is the logical relationship between StarCraft and General Intelligence? How did "StarCraft" fascinate AI giants collectively?

Before starting a detailed analysis, we must reiterate the common sense: top AI experts are not game nerds (at least not at work), and the fact that AI can play games itself has no value to them. Just like AI's achievements in Go, it is by no means intended to show off how powerful AI is or to humiliate the traditional culture of Asian countries.

AlphaGo chose Go because this intellectual sport has the characteristics of not being violently exhausted and can test AI's non-computing abilities. For example, we see that the final version of AlphaGo has demonstrated layout capabilities, judgment on abstract situations, and even creativity in specific behaviors-capabilities that traditional computers will never have.

For AI researchers, games are just a method, not a goal. AI companies have only two goals: one is to make better AI, and the other is to use it to sell money. However, games have natural laboratory attributes in training AI: supported by data, clear criteria for success or failure, and a large amount of human training data are available.

With this consensus, we can begin to crack the AI giant's plot for StarCraft. But the first step is to understand what the AI company has done to "StarCraft" and what is the purpose of doing these things?

What have DeepMind and Facebook done with StarCraft?

First of all, let's take a look at what Facebook and DeepMind released in this very "challenging" move.

In chronological order, the first is Facebook. Simply put, Facebook released a 365GB "Battle of the Mother Nest" game dataset containing more than 60,000 items of content. Of course, these data are not for people to see, but are data specifically used for machine learning task training.

What is particularly noteworthy is that Facebook specifically emphasized the versatility of these data in its published paper. They can adapt to different algorithms and different platforms, and can be said to provide relatively standardized machine learning training data.

When we come to DeepMind, things get a little complicated. Their target is not "Battle of the Mother Nest", but the more advanced "StarCraft 2". DeepMind released an integrated AI training kit called SC2LE (StarCraft 2 Learning Environment).

This gift bag contains four items, as well as a paper dedicated to the machine learning environment of StarCraft 2.

First of all, DeepMind also gave a game dataset, including anonymous game data from 65,000 games officially recovered by Blizzard. It will continue to increase in the future.

Secondly, the "StarCraft 2" machine learning programming portal developed by Blizzard was released to facilitate researchers and developers to connect their own agents to games for research.

In addition, DeepMind has also open-source its own PySC2 toolkit, making it easier for researchers to train their own agents.

Finally, the gift package also includes a series of enhanced learning mini games abstracted from StarCraft 2. These mini games allow researchers to more easily test the effects of agents in specific scenarios.

In this way, DeepMind and Blizzard scientists were more considerate. They not only delivered the staple food, but also included various tableware and desserts, as well as food instructions.

But both DeepMind and Facebook have only one goal: to provide the most convenient and comfortable research environment possible and attract more developers to join the cutting-edge AI training of StarCraft.

The basic motivation for this is that the actions and scenarios of games like "StarCraft" are almost endless, and it is impossible for a company's manpower to carry out comprehensive in-depth development and experimentation. Therefore, open source and sharing data, helping more researchers skip the basic steps and directly study cutting-edge and specific action algorithms and multi-task coordination solutions, is the real focus of AI companies at present.

This action idea comes from the particularity of real-time strategy games. Here we can begin to answer another question: Why does it have to be "StarCraft"?

Why must it be StarCraft?

To train AI systems such as deep learning, there are three most used games: mini games, sandbox games, and real-time strategy games. But among these three, AI in mini games and sandbox games plays a single agent. Only real-time strategy games provide unique training characteristics: complexity and collaboration.

Just now OpenAI's AI defeated the top DOTA2 players, Musk excitedly tweeted to celebrate (he didn't forget to mention the AI threat theory by the way), but many scientists from other AI giants were quite disdainful. Why?

The reason is that AI in DOTA2 in 1V1 mode is only suitable for one agent, the target is relatively single, and a 1-on-1 encounter is not an incomplete information game. Human speed and responsiveness are definitely not comparable to AI. The value of OpenAI's AI lies more in using the system to independently learn the rules of e-sports.

Large-scale real-time strategy games have completely different environments:

1. Complex and changeable environments test the agent's ability to understand the changes in space, time and data of a large amount of environmental information.

2. Cooperation from many independent units. The technology that human players call "micro-manipulation" is to test the ability of multiple units, buildings, and formations to operate collaboratively in a melee. This is a core test for AI.

3. Game of incomplete information. This kind of game starts with the fog of war and it is impossible to observe the movements of opponents. Agents are required to make layout and long-term judgment.

These characteristics make real-time strategy games one of the most complex AI experimental environments known, knowing that the goal of AI is not to win, but to win.

As for why it must be "StarCraft", there may be several reasons:

First, Blizzard has the willingness to create AI through open source, and the materials and interfaces of "StarCraft" itself are also very smooth. The cost of digitization of the entire game is very low.

Secondly, compared to real-time strategy games that focus on effects and graphics,"StarCraft" has stronger competitive attributes. Its actions are numerous and elements are complex. Like Go, it has basic characteristics that are difficult to be disassembled by violent calculations.

At the same time, due to its strong competitive attributes and long history of competition,"StarCraft" has a very rich discussion on tactics and strategies, and the value of each sub-action basically has a basis for judgment, which is one of the prerequisites for a machine learning system.

In addition to these three points, there is actually the most important reason: "StarCraft" is the real-time strategy game with the largest number of battles, and the construction of platforms such as Battle. com is very complete-in other words, there is enough training data left for AI.

The real purpose of AI giants

would rather adopt an open source crowdsourcing model than overcome the ultra-complex training environment of "StarCraft". Are AI giants satisfied? What do they want to get from it?

Reverse based on DeepMind's principle of disassembling "StarCraft 2" into a series of Mini games. We can know that what AI companies hope is to crack one detailed action after another and put them together to form a large collection of agents. And this collection will bring together not only countless solutions, but also the common capabilities underlying these solutions.

If we compare it with Go, we can infer at least four abilities that Go cannot give from StarCraft as a training environment:

1. Machine memory. Different from board games, past information in real-time strategy games may be completely erased, such as the soldiers just created died... but this information will influence the next story. This requires AI to have memory and use memory to provide countermeasures. Rapid adjustment based on memory will be a new and commercially valuable AI capability.

2. Long-term planning capabilities in weak information environments: As mentioned above, games like StarCraft are completely information-closed at the beginning. What you do at the beginning may have nothing to do with the outcome of the war, but it has a real causal relationship. This kind of human unique long-term planning and adjustment planning ability corresponds to the machine's prediction and judgment ability.

3. Multi-agent collaboration capabilities. Through a keyboard and mouse similar to humans, how an agent can direct a large number of agents to collaborate on the terminal is definitely a fascinating topic. Even strategic sacrifices, setting up bait and concentrating firepower may be compared to the central role of AI in real society in the future.

4. Coherence of movements: Everyone who plays games knows that the key to winning is to play the "rhythm". The so-called rhythm comes from how each command of the player connects and whether it is consistent. The same is true for AI. It is not difficult to surpass human wisdom in one detail, but how to connect every action and maximize the overall value is the key to AI's forward development.

The most exciting part of these four directions is that they all correspond to the abilities of people in reality-not only the intelligence of humans to recognize and interpret the world, but also the "kinetic energy" of humans to react on the physical world to remember, collaborate, and persist.

At this point, perhaps we can understand why StarCraft is called the key to general intelligence. Because it foreshadows the possibility that AI will learn to be more similar to the human mind in a more chaotic and real environment.

It is not only a transition from an experimental environment to a real environment. In the long run, such games will be completely conquered by AI and may even become the key to the transition from agents to humanoid agents.

Even if not to mention such grand propositions, similar agents may become the cornerstone of huge business changes such as AI replacing stock analysts, advertising planners, and lawyers-at least this kind of smart game may not be as difficult as playing "StarCraft" with top players.

AI company's love for "StarCraft" is now understandable.

Editor: Mary