AGI technology and making videogame characters alive

This is an article version of Mind Simulation CRO Vasily Mazin speech for cancelled AI Everything 2020.

This article is dedicated to Artificial General Intelligence. And the role videogames play in the process of its creation. There are going to be some difficult concepts and terms in my talk – however it wouldn`t affect the overall understanding in any way.
Sure, nowadays Narrow AI is huge, popular thing — voice recognition, speech synthesis, video editing, budgeting and forecasting software, et cetera are everywhere. However more and more people start talking about Artificial General Intelligence.
To start things off, we need to understand what AGI is. Artificial General Intelligence can solve all tasks and problems that people can. Let's look at some definitions of AGI, which given by AI researchers.
  • “General intelligence involves the ability to achieve a variety of goals, and carry out a variety of tasks, in a variety of different contexts and environments”

    Ben Goertzel
  • “The essence of intelligence is the ability to adapt with insufficient knowledge and resources”

    Pei Wang
  • “Intelligence measures an agent's general ability to achieve goals in a wide range of environments”

    Marcus Hutter, Shane Legg
    As you can see, all those definitions are pretty similar, only small details vary. Now let's try to formulate the most general definition of AGI.
    • “AGI is the ability of a system to solve any tasks and problems in difficult enviroments, even if its recourses are limited.”

      This definition is super broad and needs further clarification: What are any tasks? What kind of difficult environments and resources are there? How limited are those resources?
      However it is not that easy. Nowadays, the development of Artificial Intelligence still faces some problems. The most important ones are listed below.

      Not achieved milestones in Artificial Intelligence

      Explainable AI

      Explainable AI allows for human beings to understand the reasons why AI makes desicions. It is necessary to do verification, AI rework and AI fine-tuning.

      Transfer learning

      Transfer learning is the possibillity of transferring knowledge between different AI systems, as well as between man and machine so that the knowledge accumulated by an individual system or a human expert can be fed into a different system for use and fine-tuning.

      One iteration learning

      Few-shot learning. Systems need to be developed that can learn with the use of a small amount of materials, in contrast to the current deep-learning systems, which require massive amounts of specifically prepared learning materials.

      Structured prediction and learning

      Structured prediction and learning is developing learning technologies based on the representation of learning objects as multi-layered hierarchical structures, with lower-level elements defining higher level ones. This could prove an alternative solution to the problems of fast learning and strong generalization.

      Catastrophic forgetting

      Solving the problem of catastrophic forgetting, which is pertinent to the majority of existing systems: a system originally trained with the use of one class of object and then additionally trained to recognize a new class of objects loses the ability to recognize objects of the original class.

      Incremental learning ability

      Achieving an incremental learning ability, which implies a system’s ability to gradually accumulate knowledge and perfect its skills without losing the previously obtained knowledge, but rather obtaining new knowledge.
      As you can see from this list, many problems only exist because of neural nets and deep learning. However those are not the only approach to AGI. With the use of hybrid approaches and different building intelligent systems methodics, we have turned those problems to our advantage if we use hybrid approach and combine different techniques of building intellectual systems together. We have been trying to solve this difficult scientific and engineering challenge for ten years. Philosophers, psychologists, linguists have been helping us during the whole process. We also have been tracking new achievements and findings in neuroscience research. The functional level in the brains is especially intriguing to us. So even if there's still a lot of work to be done, we already turned this list of problems into a list of technology advantages and can show some very big accomplishments.

      Tip!

      You can read more about the AI core and its benefits on the technology page.
      There are still two interesting challenges:
      - how to check the fact that AGI is achieved?
      - how to measure how close is achievement of AGI?

      How to check the fact that AGI is achieved?

      The problem is there`s no clear definition of intelligence. It makes the development of theoretical and emperical aspects of testing and evaluation difficult. The same thing is with making it fit the requirements of AGI. Even the popular Turing test is not perfect and has some serious drawbacks since it is anthropocentric and does not have any ways to check the actual levels of intelligence.
      Fortunately, there`s many ways to test AGI — some of them are rather amusing, some are quite serious. Let`s have a look at them.
      • “A machine / the AI enrolls in a university, taking and passing the same classes that students would, and obtaining a degree.”

        The Robot College Student Test (Ben Goertzel).
      • “Steve Wozniak has suggested to create a machine which is required to enter an average American home, find the coffee machine, find the coffee and the mug, add water and brew the coffee. ”

        The Coffee Test. Steve Wozniak
      • “A machine works an economically important job, performing at least as well as humans in the same job.”

        The Employment Test of Nils John Nilsson.
        Another criterion for achieving AGI can be the solution of various AI-complete problems. Such tasks can include: natural language understanding, computer vision, Bongard problem solution, scientific articles review. So basically any complex problems that include unforeseen circumstances and require the level of AGI. Even fully self-driving car is an AI-complete problem. So we are not going to see completely self-driving cars, that can deal with any road situation, any time soon.

        An even more difficult question is how to measure AGI?

        Much attention has been recently focused on such AGI development metrics as the Universal Measure of Intelligence. Researchers Shane Legg, Marcus Hutter, Jose Hernandez-Orallo and others have contributed greatly to its development.
        This metric is based on the following unofficial definition of intelligence: "Intelligence measures an agent's ability to achieve goals in a wide range of environments".
        This theoretical measure of Intelligence has a number of positive qualities to it. It can evaluate both simple and universal agents, intelligently organize agents. It is also a continuous measure of intelligence, and is not anthropocentric. But there is also a big problem - it is just a theoretical definition and is not suitable for direct evaluation of real agents. Nowadays, we have test systems based on an extended version of the Universal Measure of Intelligence, but there is still a lot of work to be done.

        Still, how is it possible to connect video games and virtual reality with the task of creating AGI?

        Video game worlds are mini copies of our world with some changes and simplifications. Characters, who are actually agents, can solve various tasks in them. Moreoverб video games computing resources and amounts of knowledge are very tightly limited.
        As you can see, all of this comply with the AGI definition that I`ve mentioned.
        AGI and games can enrich each other. On the one hand, a game is an environment for the development of AGI. On the other hand, with the help of AGI, you can bring to life so-called open worlds that try to give freedom of choice to the player. Such games are very difficult and expensive to create, because both freedom and non-linearity require a huge amount of resources to develop additional content, that some players may not even see.
        In these games, doing side quests might lead to a gradual loss of interest if players do not know how to entertain themselves. After all, the world is essentially dead, and game characters use only pre-written lines.
        This is what it looks like. The yellow line is the main plot, and the gray one is just some extra information . We can't ask our own questions or talk to the characters. To start a dialogue with most characters is not even an option.
        When the main quest ends, everything stops moving and becomes just a set. However players love to impact the fate of the world or charachters, which makes non-linearity very important. Video game developers know that.
        • “The player wants to affect the world, and what happens in the story. Many games strive to find ways to give this power to the player, but the fact is that we just can’t afford to do it on a huge scale, in most cases, while still having anything resembling a crafted story. The cost of developing entire game areas, missions, characters, and so on, which most players will never see because they took a different path through the story, is just too prohibitive.”

          Evan Skolnick. Scriptwriter of Dying Light и Mafia III
          Is it possible now? Is it possible to fill the video game world with thousands of digital personalities? Yes, it's difficult, but possible.
          To conduct an experiment and to demonstrate the possibilities, we have decided to take the world of The Witcher 3: Wild Hunt video game, and to revive its main character Geralt of Rivia and some of his friends. To create their digital personalities that will behave like real people, have unique character traits, a body of knowledge, and so on. And we have done it. The first results of our work can be checked out at our booth. We have brought Geralt here. I would also like to point out that all of it works completely offline on an ordinary laptop.
          Why is it important for the industry? Because creating advanced and well-written dialogues for characters takes lots of time, resources and money. However, the usage of agents with AGI technology allows to create thousands of different personalities on the basis of one in no time and with minimal costs.
          This tech will change people’s minds about video games and AI in general, because games are a launching pad for AI.
          Museum guides, personal assistants like Jarvis, autonomous robots – everyone can be a digital personality.
          We would like to finish this article with some quotes that may make you smile and think.
          • “In the future the machines will become self-educated. They will watch the world as children do.”

            Yann Lecun
          • “The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.”

            Edsger W. Dijkstra
          • “Computers are getting smarter all the time. Scientists tell us that soon they will be able to talk to us. (And by 'they', I mean 'computers'. I doubt scientists will ever be able to talk to us.)”

            Dave Barry