Artificial General Intelligence and General Game-Playing Agents

Posted bykriskstoyanovPosted inWeek 3

Introduction:

Throughout the years, advancements in technology have pushed the limits of what can be achieved with AI. As computational power grew the overhead of developing advanced behavior for achieving complex tasks has decreased but an open challenge is still present in the task of developing an artificial general intelligence (AGI). The main concept behind it is the inception of lifelike AI which has the ability to understand and learn any intellectual task by itself and can influence its cognitive abilities and discern its own and others’ behavior.

General Video Game AI:

Games are a popular testbed for AI benchmarks. Research in the field has enabled some interesting advances in algorithmic AI, such as the use of parallelized Alpha-Beta pruning and Monte Carlo Tree Search in traditional board games like Chess and Go. Designing the behavior of these agents through the controllers which implement these algorithms involve training them under highly specialized conditions which inherently prevents them from being efficiently transferable to a generalized space. General Video Game Playing (GVGP) is a sub-domain of Game AI which aims to provide a framework defined in Video Game Description Language (VGDL) for creating generic enough agents that are capable of playing any given game without pre-computed game-specific heuristics, in possibly unknown environments.

An agent implemented in such environments must be able to select moves in real-time, providing a valid action in no more than 40 ms at each time step. The controller it uses receives information about the game state and it is its responsibility to discover the game mechanics without being aware of the victory conditions. The agent is provided with a tool, in the form of a forward model, which allows it to simulate its actions and roll the game forward to one of the next possible states which allows it to reason about the environment. The forward model is very fast and almost all successful agents simulate hundreds or thousands of game states for each decision taken.

Automatic Game Design:

The GVGAI framework is expanded into multiple tracks and can be implemented as a toolkit for game, level and rule generation and use the player experience to influence its behavior. This concept became an active research topic in the late 2000’s where development work was first done in the area.

  • Game Generation: Provide AI controllers which automatically generate new games or game instances by tuning game parameters. Work has been done on this concept by providing a particular theme, a database of game objects, or searching spaces of game rules, with which the participants can generate new games. Automatic tuning is achieved using search-based and population-based methods that have been applied to game parameter optimization aiming at maximizing the depth of game variants or finding more playable games.
  • Multi-Player GVGAI: The multi-agent game domain is a popular development field with competitions being held since 2011. The interface for the two-player planning track was initially developed for two or more players, so it has the potential to be expanded to a multi-player planning track in which an agent is allowed to control more than one player or each of the players are controlled by a separate agent. This future track can be expanded as a multi-agent learning framework, providing a two-or-more-player learning track.
  • Turing Test GVGAI: Determining if an agent that is playing a game is a human or a bot is a challenge that has been subject of study for many years, and the idea of applying it to a general video game setting is not new. This concept offers an interesting opportunity to extend the framework to having a Turing Test Track where participants create AI agents that play like humans for any game that is given and an audience discerns whether an agent or human is playing the game and what differentiating factors exist.
Further Development:

The GVGAI framework offers the most comprehensive system to date for evaluating the performance of general video game playing agents, and for testing general purpose algorithms for creating new games or new content for novel games. The framework has been used in multiple international competitions, and has been used to evaluate the performance of hundreds of general video game agents.

There are several improvements and additions to the framework that can be done and would potentially affect all existent and future competition tracks. One of these continuous modifications is the constant enlargement of the games library. More games are added with each subsequent competition while improvements in the GVGAI functionality are made which has the potential to create infinite number of games that can be integrated into the framework.

Adding more games can also be complemented with compatibility with other systems. Other general frameworks like OpenAI Gym, Arcade Learning Environment (ALE) or Microsoft Malmö count on a great number of single or multi-player, model-free or model-based tasks. Intefacing with these systems would greatly increase the number of available games which all GVGAI agents could play via a common API. This would also open the framework to 3D games, an important section of the environments the current benchmark does not cover.

With regards to the agents, another possibility is to provide them with a wider range of available actions which they can carry out simultaneously or be used to form a continuous action space. This would enhance the number of legal combinations for the agent to choose from at each decision step.

The agent tracks cater for planning agents able to exploit a fast forward model, and learning agents that must learn to react sensibly without the benefits of a forward model. The planning track already comes in single- and two-player versions, while the learning track is currently single-player only, but with two-player version envisaged. Long term learning may also be used within the planning track as successes in Go indicate what can be achieved by combining learning and planning.

Challenges and Opportunities:

AGI and GVGAI are plausible now more than ever before due to the exponentially higher amount of compute resources and data being available but they still remain largely the topic of continued research and development with the help of giants like Google and other major investors pushing the frontier through smaller companies like DeepMind. Major scientific and technological breakthroughs are necessary to overcome the current widespread adoption of “narrow-AI”, which performs reliably in a single domain, and transition to a more general model that can be adopted across multiple sectors. If researchers one day succeed in building a human-level AGI, it will probably include expert systems, natural language processing and machine vision as well as mimicking cognitive functions that we today associate with a human mind like learning, reasoning, problem solving, and self-correction. However, the underlying mechanisms may differ considerably from those happening in the human brain just as the workings of today’s airplanes differ from those of birds.

Credits:

References:

Published by kriskstoyanov

Third Year Undergraduate Student

Leave a comment

Design a site like this with WordPress.com
Get started