Measuring AI Advancement Through Games in 2024

Games have served as milestones for showcasing progress in artificial intelligence for decades. As AI capabilities improve, conquering games that were previously considered impossible for computers demonstrates concrete advancement. Some of the most notable achievements in the history of AI involved game-playing systems defeating top human experts in challenging environments. Games provide controlled yet difficult challenges that push AI capabilities forward.

This article will explore the relationship between games and AI, looking at major breakthroughs over the years and how games have been instrumental for advancing machine intelligence. It will also discuss current trends and the future outlook for game-playing systems as measures of AI progress.

A Look Back at AI‘s Game Victories

Game-playing has been a part of AI research since the earliest days of the field in the 1950s. But it wasn‘t until the late 1990s that an AI system achieved a major public victory against human opposition.

May 1997 – IBM‘s Deep Blue supercomputer defeated world chess champion Garry Kasparov in a 6-game match, becoming the first computer system to beat a world champion in a classical chess match format. Deep Blue calculated up to 200 million chess positions per second using specialized chess hardware. It combined this brute computational force with chess knowledge encoded by grandmaster-level programmers.

February 2011 – IBM achieved another milestone when its Watson system beat two former champions on the Jeopardy! game show quiz competition. Watson processed natural language questions and answered correctly in under 3 seconds with over 90% accuracy in the two-game match. This demonstrated an ability to understand nuanced linguistic clues and quickly query its vast knowledge database.

March 2016 – The ancient Chinese board game Go represented a grand challenge for AI due to its complexity, with more possible board configurations than atoms in the universe. Google‘s DeepMind AlphaGo program beat Lee Sedol, one of the world‘s top professional Go players, by a score of 4 games to 1. AlphaGo used neural networks trained by reinforcement learning from self-play, rather than brute search, to develop its strategies.

January 2017 – Libratus, an AI system from Carnegie Mellon University, defeated four of the world‘s best professional poker players at no-limit Texas hold ‘em poker. This imperfect information game requires skills like bluffing, unpredictable behavior, and opponent modeling that posed new challenges for AI. Libratus won chips worth over $1.7 million in the 120,000 hand Brains Vs. AI competition.

In addition to classic board games and poker, AI systems have attained superhuman performance in videogames as well. In 2015, Google‘s DeepMind created an AI agent that could learn to play Atari 2600 games better than humans just from pixel inputs, setting records on over half of the games it tried. DeepMind‘s AlphaStar system achieved Grandmaster level performance in StarCraft II in 2019, mastering a real-time strategy game with partial visibility of the game state.

These examples demonstrate how AI capabilities have advanced to surpass the best human players across a variety of games, from turn-based board games like chess to real-time video games like StarCraft II. Games provide concrete benchmarks for measuring AI performance against the high bar of human expertise.

Why Games Make Good AI Benchmarks

Games offer controlled environments with fixed rules for testing an AI agent‘s abilities on cognitively demanding tasks. They provide instant feedback on the results of actions through scores and win/loss outcomes. This makes it straightforward to compare an AI agent against human performance. The rules are unambiguous, so an AI can‘t cheat or exploit loopholes to appear more capable than it really is.

Many games also have logical separability – skills developed on one game often transfer to others. For example, learning heuristics that improved chess play also enabled better checkers and Othello play without much specific adaptation. So progress on conquesting one game often leads to breakthroughs on others as well.

Games are recognized as difficult tests of intelligence. They require skills like strategic planning, opponent modeling, dealing with uncertainty, and quick pattern recognition. They offer a microcosm for studying an AI‘s proficiency on tasks that resemble human reasoning in many ways. Outperforming expert humans in challenging games is a persuasive demonstration that AI capabilities are reaching or exceeding human levels.

While games are simplified environments compared to the real world, they provide challenging stepping stones on the path toward artificial general intelligence. As research exposes limitations of AI agents in games, it inspires new methods to overcome these weaknesses and expand capabilities.

The Rise of Video Games for AI Research

In recent years, increased focus has shifted to testing AI on real-time video games. Unlike turn-based board games, video games require quick reactions and handling lots of data on the fly as the environment rapidly changes. Partial observability is also common, where players lack full information about the game state. Multiplayer games add further complexity through factors like team coordination and opponent modeling.

Real-time strategy (RTS) games pose particularly tough challenges. In addition to quick decisions, RTS games involve:

• Management of resources and long-term planning
• Control of dozens to hundreds of units
• Dealing with hidden information
• Adapting strategy to opponent‘s actions
• Vast action spaces, with >10^26 possible options per move

In 2016, DeepMind‘s AlphaGo receiving widespread acclaim for mastering Go demonstrated AI‘s potential on complex board games. This inspired greater enthusiasm around tackling difficult video games next.

OpenAI began testing bots on 1v1 gameplay in Dota 2, a multiplayer online battle arena game with two 5-player teams. While their bot fell short of elite human players in a full 5v5 match, it exhibited complex behaviors like sacrificing pieces to gain long-term strategic advantage. Dota 2 has >10^1,000 possible states, presenting a massive challenge for AI.

DeepMind partnered with Blizzard Entertainment to release the full StarCraft II game as an AI research environment in 2019. StarCraft II is considered an AI "grand challenge" due to factors like real-time play, imperfect information, and massive action spaces. Top human players demonstrate complex skills like strategic thinking, adaptability, and spatial reasoning in this game.

RTS games require a more general intelligence beyond brute computation, which is why they are such difficult environments for AI systems today. They inspire developing AI that can handle real world complexities like managing uncertainty, adapting dynamically, and coordinating with teammates.

AI that Can Create Games

Beyond playing existing games, AI has shown emerging potential for designing and generating novel games. Some examples:

• Researchers at Georgia Institute of Technology developed an AI agent called GameGAN that learned to generate complete functional levels after watching just a few minutes of gameplay video for Super Mario Bros. The AI created levels that were rated as highly playable by humans.
• Unity, creator of one of the most popular game development tools, is using AI to develop helper systems aimed at democratizing game creation. For example, their Obstacle Generator can automatically spawn obstacle courses for 3D platformer games.
• OpenAI has conducted research on using AI to generate 3D game assets like trees and furniture. Their Kate model can produce diverse, realistic items controllable via text prompts.
• Ubisoft is working on AI tools like the Map Generator that can automatically create reusable, high-quality game levels to accelerate worldbuilding.

These examples demonstrate how AI is moving beyond just playing games designed by humans toward actually creating novel, playable games from scratch. This requires a deeper understanding of what makes games fun and designing content optimized for gameplay.

Games as Stepping Stones to General Intelligence

While games provide useful testbeds for developing AI, a system that achieves superhuman performance in games may still fail at many real world tasks. Games only indicate proficiency within a narrow domain. True intelligence involves excelling at a wide diversity of cognitive challenges, not just games.

Chess Grandmaster Garry Kasparov, who lost to Deep Blue in 1997, has emphasized this point. After his defeat he said, "What computers are good at is crunching numbers and using brute force. Humans are good at thinking and using intuition." Games like chess may appear complex, but provide limited tests for general intelligence.

So while games can produce AI improvements, the techniques developed for gaming must ultimately be adapted and combined to achieve artificial general intelligence (AGI) – the broad, flexible intelligence of humans. Self-play and simulations can even lead to AI agents that overfit on a particular game and falter when conditions change.

To directly evaluate real world competence, we need AI benchmarks focused on how agents behave in open-ended, interactive environments requiring common sense and sound judgment. For example:

• AI Safety Gridworlds – Assess how AI agents behave in dangerous scenarios with humans in the environment.
• AI Habitat – Physics simulator for household chores, providing interactive environments for developing AI assistants.

Ultimately, measuring AI advancement requires a diverse toolkit of games, simulations, and real world tasks. Games provide a microcosm for studying narrow abilities, while applied benchmarks offer a more realistic assessment of general intelligence needed for the complexities of the real world.

The Future of Games for Advancing AI

It‘s clear games have pushed AI progress through the concrete challenges they provide. What does the future hold in this dynamic area of research?

• Open-ended games like Minecraft and Roblox that support limitless emerging complexity will stretch AI capabilities and creativity.
• Massive multiplayer worlds like Fortnite will require AI to handle huge state spaces and unpredictable human behavior.
• Realistic VR environments like activeworlds will demand more human-like visual processing and visuomotor control from AI.
• General Game Playing competitions like the GGP Competition motivate developing AI that can excel at any game with minimal adaptation.

As algorithms and compute power improve, conquering more complex games indicates steady progress toward human-level skill. But games are ultimately just simplified facsimiles of reality. The true test will be how well game-playing AI can transfer its abilities to real world tasks. Can a Dota 2 playing bot also drive a car, shop at a grocery store, or hold a conversation as well as a human?

Through games, AI will keep extending what machines can do better than the most skilled humans. While games provide sandboxes for testing narrow abilities, the final score that matters most is how AI research advances real world competence. On that front, we‘re still early in the match. But game-playing systems demonstrate how AI capabilities are steadily leveling up.