When Deep Blue defeated Garry Kasparov in 1991, it was a triumph of engineering, but it exposed a fundamental flaw in artificial intelligence: brittleness. Deep Blue was a master of chess, but it could not play a single hand of poker or a game of tic-tac-toe. As argued by Levinson et al. in The Role of Chess in Artificial Intelligence Research, brute-force engineering for a single domain fails to teach us about true intelligence. Decades later, even as deep reinforcement learning conquered Go and StarCraft, this core limitation remained. As Yin et al. observe in AI in Human-computer Gaming: Techniques, Challenges and Opportunities, modern high-performance agents are still largely restricted to specific maps or environments, failing to generalize when the context shifts.
To build human-like intelligence, AI must move away from specialized systems and toward General Game Playing (GGP) and General Video Game Playing (GVGP). The goal here is simple yet profound: design a single agent that can accept the rules and state of an entirely unknown game and play it competently without any human intervention or game-specific reprogramming.
The Core Challenge: Decoupling Strategy from Rules
In a standard reinforcement learning setup, the agent's architecture is hardcoded to match the game. The neural network expects a specific input shape (e.g., an $8 \times 8$ chess board) and maps it to a fixed set of actions.
In general game playing, this assumption is shattered. As Banerjee et al. explain in General Game Learning Using Knowledge Transfer, a generalist agent must operate in environments where rules, objectives, and state spaces are completely arbitrary and unknown beforehand. The agent cannot rely on a fixed board geometry or pre-baked heuristics.
To bridge this gap, researchers have focused on creating universal game description frameworks. In A Practical Introduction to the Ludii General Game System, Browne et al. introduce a "ludemic" approach. Instead of coding games from scratch, games are expressed via a high-level grammar of fundamental units called ludemes (e.g., "tiling", "step", "capture"). By decomposing complex games into these universal building blocks, a single engine can interpret and run thousands of distinct board games, giving the AI a standardized interface to reason about mechanics.
[Game Description Language / Ludemes]
│
▼
[Universal Simulator / API Wrapper] ──► [Generalist Agent]
▲
│
[Raw Visuals / Pixel Inputs]
When transitioning from board games to video games, the representation problem shifts from symbolic logic to continuous perception. As Levine et al. detail in General Video Game Playing, video games introduce real-time constraints, complex physics, and hidden information. An agent cannot wait to parse a complete game tree; it must act immediately based only on the current state and available actions, forcing the underlying algorithms to find abstract strategic commonalities across visually diverse worlds.
Searching and Learning Without Heuristics
If an agent does not know the rules of a game beforehand, how can it evaluate whether a move is good or bad? Traditional AI relies on hand-crafted evaluation functions (e.g., assigning point values to chess pieces). In GGP, this is impossible.
The breakthrough solution is Monte-Carlo Tree Search (MCTS). As surveyed by Świechowski et al. in Recent Advances in General Game Playing, MCTS is uniquely suited for general gaming because it does not require domain-specific knowledge. Instead, it evaluates a state by running random or semi-guided simulations (rollouts) into the future to see which paths naturally lead to victory.
However, vanilla MCTS is often too slow for real-time video games. To make it work in dynamic settings, researchers must optimize how the search tree is explored under strict time budgets:
- Dynamic Rollouts: In GreedyUCB1 based Monte-Carlo Tree Search for General Video Game Playing Artificial Intelligence, Park et al. introduced a modified selection policy and game-analysis rollouts. By analyzing the immediate consequences of actions during the game, the agent dynamically biases its search toward promising branches.
- Search Enhancements: In Enhancements for real-time Monte-Carlo Tree Search in General Video Game Playing, Soemers et al. proposed a suite of optimizations that allow MCTS to better allocate its limited computational resources, dramatically increasing win rates across dozens of unknown real-time games.
From Pure Search to Foundation Models
While search works well when a simulator is available, another branch of the field focuses on learning visual and structural regularities directly from pixels. Early approaches, such as the one proposed by Hausknecht et al. in HyperNEAT-GGP: a hyperNEAT-based atari general game player, used neuroevolution to exploit the geometric regularities of game screens, mapping visual spatial patterns directly to actions across different Atari titles.
Today, this concept has evolved into large-scale pre-training. In NitroGen: An Open Foundation Model for Generalist Gaming Agents, Magne et al. demonstrated that by training a foundation model on 40,000 hours of gameplay across more than 1,000 distinct games, the agent develops an intrinsic understanding of gaming concepts (like gravity, enemies, and keys). When dropped into an entirely unseen game, the model does not need to start from scratch; its pre-trained representations allow it to generalize and adapt almost instantly.
Measuring True Generality
How do we prove that an agent has actually developed general problem-solving skills rather than just memorizing a specific set of challenges? As Kersting notes in Making AI Smarter, complex strategy games serve as the ultimate testbed for AI because they force systems to develop procedural reasoning and flexible human-computer interaction that go far beyond simple pattern recognition.
To evaluate this, the community relies on massive, multi-game benchmarks and competitions where agents are tested on hidden, unseen environments:
- Massive-Scale Testing: In The 2022 Ludii AI competition, Piette et al. describe an evaluation setup where agents were forced to play over 1,000 different board games. This extreme variety ensures that any hardcoded, game-specific logic will fail, leaving only truly general algorithms to succeed.
- The Structure of Machine Generality: By analyzing performance data from these competitions, researchers can study the nature of artificial intelligence itself. In AI Generality and Spearman's Law of Diminishing Returns, Hernandez-Orallo utilized data from the General Video Game AI (GVGAI) competition to show that AI generality behaves fundamentally differently than human intelligence, failing to follow classic psychological patterns like Spearman's law.
- Efficient Benchmarking: Because running an agent across hundreds of games is computationally staggering, Stephenson et al. developed an information-theoretic framework in A Continuous Information Gain Measure to Find the Most Discriminatory Problems for AI Benchmarking. They proved that by using a mathematically optimized subset of just a few highly informative games, researchers can evaluate an agent's broad strategic generality just as accurately as testing on a massive suite.
Ultimately, general game playing works because it forces a separation of concerns: the environment provides the arbitrary rules, while the agent provides a flexible, domain-agnostic engine for search and representation. By mastering the ability to learn how to learn, these systems bring us one step closer to adaptable, general-purpose artificial intelligence.
Key Foundational Work
The Role of Chess in Artificial Intelligence Research (1991) — contributed 1 concept elements to this direction.
General Game Learning Using Knowledge Transfer (2007) — contributed 1 concept elements to this direction.