Bright! FRONTEO Official Blog

Bright! FRONTEO Official Blog

Providing and communicating with humans with AI experiences


≪Table of Contents≫
1. "Game AI is the elusive thing to study" Mr. Miyake
2. Symbolism and connectionism
3. "My guess is that it will happen somewhere until 2030," Miyake
4. What physical expression and game AI have in common
5. "Ideal has a fundamental motion generation function, and I want to put individual motions on it." Mr. Miyake
6. Game AI output
7. Communication between people and AI

"Game AI is the elusive thing to study" Mr. Miyake

FRONTEO CTO Hideki Takeda (hereinafter referred to as Takeda) Recently, Mr. Miyake's treatise * has become a hot topic.The history of game AI is comprehensively written, and when I read that, I thought it was interesting to understand that the flow of game AI was like reading a history book.

Yoichiro Miyake, Specially Appointed Professor, Graduate School of Artificial Intelligence Science, Rikkyo University (hereinafter referred to as Miyake) There are not so many papers in the field of digital games from industry, so I include an overview of this field as an explanation of the prerequisite knowledge.A treatise is usually a narrower search, but I think it is a field that many people are not familiar with, so I carefully explain the overall picture of this field before entering this paper.

(I.e. Certainly, many of them have a limited range.

(I.e. Yes.However, digital game AI is a field that has not yet been developed, so as a result of writing it while organizing it, it became a comprehensive content.The reason for receiving the award was that "the concept is organized so that readers in fields unfamiliar with the game can easily understand it."I think that it is expected to contribute in the future.Digital game AI is the elusive thing to study so much, and it's hard to tell from the outside what you're doing in the game industry.The information was chaotic even inside, and when I entered the game industry 15 years ago, I felt like I was in the fog and didn't know anything.So, this time, I wanted to convey that there is such a field and there is such a theory by making a general theory at the same time as summarizing it and making an example applied to it.In addition to that, I hope to convey that Miyake thinks this is a general theory.There are various research themes, but not many people want to do digital game AI.There is also a reason that there are many people who want to do it, but few people can teach it.Little research has been done at Japanese universities, except for a few laboratories.Go, shogi, and chess have clear rules, so it's easy to study, but you have to make your own digital games or bring a commercial game.Digital games can be made like that, but it is difficult to make the real thing, so academically it takes a lot of power from the beginning to proceed with research.

Yoichiro Miyake, Specially Appointed Professor, Graduate School of Artificial Intelligence Science, Rikkyo University

(I.e. I feel surprising.There are a lot of players, right?

(I.e. You're right.

(I.e. Imagine from that, I think there are a fair number of people doing research.Not really.

(I.e. In fact, there are quite a few in the world, and the Nordic countries are one of the bases.In the Netherlands and Sweden, mainly in Denmark, there are universities with research institutes for game AI.The Game Institute of IT University of Copenhagen is famous.Another base is North America.The game industry itself is thriving in San Francisco, and we are collaborating with the University of Southern California and others, and Stanford University has been doing game research from an early stage.On the East Coast, MIT was the earliest to start researching digital games in the 90's.While digital games are treated as a subculture in Japan, they are classified as computer science and art overseas, so their positions are quite different. In the early days of game AI around 1995, it was treated as a "video game" overseas, but after the appearance of 3D games, it was not just a pico-pico, but an advanced computer simulation space. The trend to study game AI in the context of computer science has taken place around MIT.In Japan, there was once a strong awareness of "Is video game research useful in the world?"However, people who were children in the 80s when video games began to appear have grown up and still play video games, so their awareness of games has changed little by little.I think that this award also reflects the social situation in which such digital games have been accepted in the world and academics.Nowadays, even the brilliant Go and Shogi AI research has had a strong wind.

* "General system and implementation of artificial intelligence in large-scale digital games-based on the actual example of FINAL FANTASY XV-"
Received the "2021 JSAI Best Paper Award 6" from the Japanese Society for Artificial Intelligence on June 21, 2020

Symbolism and connectionism

 Speaking of Go, most recently, the third AI boom, Go AI "AlphaGo," has become a major turning point.I have something to think about what Demis Hassabis (CEO of DeepMind) did.Only the point of view of improving intelligence as AI, that is, defeating a strong human player using deep learning, is emphasized, but even if he is originally a chess player and not Go itself. You're someone who has a deep insight into board games.There is an aspect that people who know the domain are doing AI.However, although there are many articles that mention deep learning, there aren't many stories about the context of knowing the domain.I thought it was strange.

FRONTEO CTO Hideki Takeda

(I.e. I agree.There are symbolism and connectionism in artificial intelligence, but symbolism = symbolism requires humans to directly embed domain knowledge, that is, rule base, knowledge base, etc.Go is a combination of deep learning, evaluation learning, and Monte Carlo simulation, all of which are simulation technologies.There is a frame, but it automatically learns in it.Little human knowledge is included.You can do some research without knowing Go so much.Even if a person who was a Go player did research, that knowledge does not have much influence directly, but rather it is like rediscovering from AI.Hassabis was a chess player, but he is studying brain behavioral science at university and has published a famous treatise.It seems that you arrived at AI through such a career flow.

(I.e. How are symbolism and connectionism treated in game AI?

(I.e. Now, in the game industry, artificial intelligence of games is being created by embedding domain-specific knowledge as Mr. Takeda said.Most of the commercial version of game AI is that pattern.Because, unlike products and research, products are responsible, so we need to guarantee what logic we have, how we can make adjustments, and how we can improve quality.Symbolism is a mechanism that improves the quality by increasing the rules and adding domain knowledge.The creator can control the evaluation value because he knows what is inside.In the case of connectionism such as neural networks, it is a black box that cannot be customized or quality controlled because it says "Algorithm from input to output."It's a story about whether this can be shipped as a product.Research on quality assurance has not progressed at all.For academics, if you're stronger than humans, you'll get results, but in the case of products, you need to be assured that your users will enjoy it, control its quality, and debug anything.Nowadays, such quality assurance can only be achieved by symbolism, so from now on, research is needed on how to make connectionism that can be controlled and guaranteed.Moreover, non-engineers must be able to control it.If it is rule-based, it can be controlled by the data input of the designer, but in connectionism, it is necessary to judge whether the learned neural network needs improvement or sufficient quality.It has developed to that extent and can finally be applied to industry, but the fact that the needs of the research have not been communicated from industry to the academic field may also be the reason why the research has not progressed.
After "AlphaGo", it was announced that DeepMind played a game called "Starcraft 2" and OpenAI played a game called "Dota 2", each of which learned about 150 years of human learning time and beat humans. However, research is still "how to surpass human beings", and from the perspective of our industry, there is a feeling that it is not there.But that momentum pushes me.

"My guess is that it will happen somewhere until 2030," Miyake

 Is there an answer from the industrial side in the situation where it is being pushed?

(I.e. The industry side is also divided into two. Half of AI engineers say that "deep learning is useless in games," and the other half say that "you have to take the time to get used to the game industry."The latter is trying to put AI in the development process, not in the game.For example, AI for autoplay that provides quality assurance is being researched.Since it is a research to replace the tester with AI, there is no problem even if it crashes in the worst case.There is a trend to try using it in such a field first, and to introduce it first in a situation where there is plenty of memory, CPU and GPU that does not require real-time performance.It is still early to incorporate it as a commercial version.It is being carefully proceeded with reference to what is being tested in non-product areas.

(I.e. So that's it.It's interesting that the introduction is progressing toward the side closer to the player.

(I.e. It's interesting, but on the other hand, compared to the situation where people outside the game industry are actively incorporating deep learning into their games, it feels like they're left behind (laughs).

(I.e. Do you feel that Miyake wants to incorporate more?

(I.e. I think you should incorporate it.Since the needs of the industry are known only within the industry, the point of the problem shifts even if the research results are taken in from the outside.If you want to incorporate a neural network that you use in real time, I think there is a way to use the neural network as you incorporate it.In the 00s, neural networks were called black technology = black magic in the game industry, and it was said that we should not use it because we do not know how many layers to use and there is no guarantee of debugging.However, deep learning has become popular, and the threshold has become lower, so we have no choice but to use neural networks.However, it is not the case that you can enter the game if you feel relieved, and you have to take on the challenge of actively incorporating it into the game.Looking at it over a long span, I think that game AI will slowly transform from the current symbolism to connectionism, but it should change dramatically in a certain year.My guess is that it will happen somewhere by 2030.I don't know what will trigger it, but the GDC conference is a technical indicator.The game AI technology announced there may have all changed to deep learning in the year.It is also a matter of who will make the first turning point, and it will be important to win at the boundary.As with online games, game companies have a history of attracting users with their technological capabilities, so if it becomes "the AI ​​of this company is amazing," it will be possible to brand it for the next 10 years. That's right.Everyone is aiming, but the difficulty is high.It may not be so difficult if you just put it in technically, but it is difficult to create a development environment after persuading artists and game designers.

What physical expression and game AI have in common

(I.e. The story changes, but I had an image while reading Mr. Miyake's treatise.I like contemporary dance and wanted to be a choreographer when I was young.

(I.e. The changes are quite drastic (laughs).

(I.e. When I was 20, I was really aiming for it (laughs).There was a famous choreographer named William Forsythe who advocated the Binary Ballistic Ballet.What's interesting about him is that he creates a digital archive of something like a motion dictionary that pairs physical expression and language, and lets dancers learn it.I'm learning and using the "language" that I used to express my body to make dancers talk.Like jazz, it has theory but also freedom.A group dance with a clear script and a part for free conversation are combined into one work.And the story of game AI was linked in me, but the way of choreography is similar to the structure in which the character moves along the story by AI or moves depending on the conditions in the game. I thought it might be.

(I.e. What kind of unit is it?Is the movement decided for the alphabet 1?

(I.e. There are some variations, for example, alphabet to sentence.

(I.e. There are cases where the word is decided at which position, and there are cases where you combine it yourself.Certainly, it's similar to a character's unit of movement.Character AI, especially in the case of symbolism, is a discrete symbol that shows each motion, but animation is a continuous system, so it is not compatible and it is the most difficult to connect discrete and continuous. What is it?If you have a dictionary of discrete movements and you want to do it continuously, what to do with the connection, if there is an obstacle, if you can not pull the bow and try to pull it down one step, you will fall from the valley Something like that happens.How to connect continuous and discrete This has not yet been found to be a good solution.This is a mind-body problem in philosophy.

(I.e. Descartes'.

(I.e. Yes.When humans perform actions such as picking up PET bottles, humans connect the movements such as stretching their arms without permission, but this ability is actually a very advanced ability.Since we are unconsciously connected, it is difficult to break down how we are doing.It's not a matter of consciousness because it's not tied to decision making.Actions associated with consciousness can be done with symbols, but that is not the case.Ballet can also be done with scripts, but it seems that it is established because humans connect the movements during that time without permission.

(I.e. that's right.That's right.It seems that the restrictions from the body side will realize the connection without permission.

(I.e. In the case of a character, you have to make that connection.

"Ideal has a fundamental motion generation function, and I want to put individual motions on it," Miyake

 I'm really interested in how the connections are designed.

(I.e. Actually, it's a very big problem, and it is the animator that guarantees the continuity of movement, but there is a graph that blends the motion between movements.From sitting to standing, the ratio of "sitting" movements and the ratio of "standing" movements are controlled.The feeling of motion depends on that.By the way, there is a difference between overseas animation and Japanese animation in this feeling, and although overseas animation has a light impression, many Japanese animations are quick and fast.It may be called realism and Kabuki.It makes a big difference in how the user feels when interacting with the character on the controller.If you focus on reality, it will definitely slow down, but the interactivity as a game is higher for fast movements.Animation is used to guarantee the smoothness of movement, and AI is used to instruct "Please stand" and "Please sit".As expected, there are some animations that cannot be guaranteed.Even though I was climbing the ladder, even if I was instructed to "kick", I fell off the ladder.How to solve it is a difficult place.A common practice is to create a multi-layer structure.I put multiple layers between decision making and animation, but the number of layers increases to solve various problems.Each has its own individual adjustments.That's the case, but the essential solution isn't. AI defines movement by itself, isn't it?Just like the ballet I mentioned earlier.There is no motion generation function because the motion is cut vertically.There is only a connection.It seems that human beings have about four stages of motion generation functions such as "balancing," "adjusting the position," and "moving the body."Humans have the ability to create general movements.Since that cannot be done with AI or robots, motion is defined from above, but ideally there is a fundamental motion generation function, and I want to put individual motions on it.What we should aim for in a long span is to create this motion generation function.Instead of adding motion capture data called animation, AI itself must create movement.That is quite difficult, isn't it?In fact, in academics, various researches are progressing centered on SIGGRAPH, and when it is found that there is an "enemy" that this can be realized, AI will judge for itself and slide to that extent, smoothing the interaction with objects. It becomes possible to express in.This is not possible with fixed motion.It's a very interesting study.

(I.e. You say "generate motion".

(I.e. I agree.It is called motion planning, motion generation, procedural generation, etc.After all, there are places where the intelligence of AI changes depending on what the output is.If you only need to play rock-paper-scissors, you only have to think about goo, par, and choki.However, when various outputs are needed, AI has to think more accordingly.As diversity emerges in the future, AI will also have to raise the level of decision-making.

Game AI output

 What is the output is an important point in terms of what kind of functions are required of AI.In the case of a game, if you specialize in movement, motion generation will be the output, but from a more meta perspective, the game itself has a play element and the goal is that the user can enjoy it, and all of them. If is the output, it has a multi-layered structure, isn't it?What do you think is the output of game AI uniquely?

(I.e. Games also have a history, and I think they are gradually changing, but basically there are loops that users can enjoy.For example, if there is a movement called sword fight, you can enjoy it for about 1 second.It's not that it's fun to do it all the time, but there is an event to defeat the opponent after that.Then you can enjoy it for about 1 minute.Even if you keep defeating grasshoppers and enemies, you will get tired of them, so I will raise the level.If you get tired of leveling up, you can get a ship and go to the next continent.That's 30 minutes of fun.A game prepares the fun of the time hierarchy like this.The old simple game is rock-paper-scissors!In the 80s, sword fighting became possible, and eventually the concept of experience points was born, and it has evolved into a complicated game system.After that, the graphics evolved and interactive stories came out, and so on, creating fun for each time scale such as 1 second, 1 minute, 30 minutes, 3 hours, 10 hours, etc., and controlling it is AI. is.The character AI of the body movement is fun for 1 second, and the meta AI that controls the whole game controls the fun of long span, and if you think you are lost, you will lead to the exit, or if you think you are tired, a new scene I think that it is the mission of game AI to provide enjoyment according to the user by sandwiching the user from both extremes, such as preparing the experience.

(I.e. When I heard the story, I wondered how it would be compared to our AI adaptation area, but in the case of FRONTEO, we substitute and support the work of professionals such as lawyers, who read documents and make decisions. I will.For example, when a lawyer finds evidence in a trial, there are many documents that must be read and judged, so after making a certain number of judgments on a small number of documents, let AI learn the judgment.It then applies a trained model to the large amount of remaining unjudged documents to make decisions on behalf of humans.The act of reading a document is a unit of time of one document in a few minutes, but these judgments are based on the business knowledge and experience that lawyers have cultivated over many years.In addition, such AI is used for contingency investigations in the event of fraud or proceedings, and for auditing, etc., to observe in normal times whether problems that lead to fraud or proceedings occur. In other words, it is a unit such as daily check, weekly / monthly, and in case of emergency, it is a long time axis of whether or not it occurs once every few years.It is not the same structure that AI handles multi-layered events from seconds to hours like game AI, but I think that it is suggestive and interesting that such a concept of time exists in game AI. I did.

(I.e. It's a little different from time, but I think the scale hierarchy is similar.I think there is a way to support according to the scale, but in the game there is also a spatial AI (Spatial AI) that analyzes spatial information, and it is an AI that prepares knowledge according to the space.The game has a multi-layered structure of time and space, but in terms of the expert knowledge that FRONTEO's AI is targeting, I wonder if there is also a hierarchy of specialized areas.

Communication between people and AI

 I think that time is quite abstract because modern society is becoming more and more complicated and knowledge can be accumulated in a place other than the body.The biggest features of game AI are (3) real-time, (XNUMX) interactive, and (XNUMX) having a body.In particular, what it means to have a body is to continuously experience the same time and space as human beings.In other words, as long as you have a body, you are involved in the world and have no choice but to receive the passage of time.That's why game AI has an element of time.For those that involve physical movement, we will design a hierarchy of time.I'm sure that theme parks are calculated for each time class.On the other hand, most modern work is abstract, and the knowledge is deeply layered except for the concept of time.There is a hierarchical structure of knowledge for each domain, and I think AI will play a role in supporting it in individual specialized areas.I think that both game development itself and the area of ​​game developers can be passed on to AI. I think it's good to be able to make games with AI alone, and I think it should be human beings who supervise it.

(I.e. Communication methods between people and AI also vary depending on the area of ​​specialization.Will you run with or ahead?

(I.e. I think that there are two issues, and it is ideal that humans ① and AI ① interact to support humans, but I don't think that can be achieved so far. It's very difficult for AI to understand people, and what's inside AI is too different from what's inside humans. The reality is that AI must support humans in other ways. I think we have to change the definition of the word "understand humans." "Collaboration" is the best way to go.When working together, you don't necessarily have to understand everything.If humans and AI can cooperate well, such as passing necessary information in a certain task or responding proactively, it is considered that they understand the human being in the task. "Collaboration in each issue" = "Understanding human beings" is what AI will do from now on.

However, if you do so, you will have to prepare AI for each frame, which is the next challenge. This AI will be used for the job A, and this AI will be used for the job B, and it will be a multi-agent support.At home, there are cleaning robots and cooking robots alone, but I think that it is not humans who manage them, but AI that manages them.It is a diagram that there is an AI that manages the whole, there are multiple AI agents that support individual tasks, and they support to surround human beings.I think this kind of thing will take shape in the future.

(I.e. In most cases, our domain is used for the narrowed purpose of using AI for specific intellectual decisions.Areas where interactions do not have to be human.However, although the decisions made by AI are gradually being accepted in the world, is it all right with AI alone, for example, in proceedings and trials?That feeling is still deeply rooted in some parts.Then humans have to explain the results of AI.It will be necessary for AI vendors like us to have an interaction that can be explained. It is an interaction with the question, "A certain number of documents related to fraud and proceedings to be dealt with have been discovered by AI. How can the evidence be guaranteed?"For example, we explain this as statistically significant and include it in the UX of the application.Human work has a wide variety of domains, and the methods of explanation accepted by each domain differ from the common denominator, and the accepted process also differs.I think that if the application side, including AI, comes close to the conventions of the human world and gives answers, it will be a useful AI in that area.As is often said, AI is different from humans in what it is good at, so I think it is necessary for AI and humans to collaborate well.
It's almost time for me, but I can't talk about it at this time.

(I.e. You can keep talking as much as you want.

(I.e. Let's talk about this continuation at another time.

(I.e. Yes.Please.

(I.e. Thank you for today.

Articles in the same category

Commonly viewed articles

ja 日本語
Machine Translation by Google. : close x