Why is It Difficult to Make Good AI for Games? | AI 101

Summary

In this video, Tommy Thompson explains the challenges of creating compelling AI in video games, focusing on the difference between AI trained outside the game (like AlphaStar) and in-game AI built during development. He explores core concepts like states, actions, and state spaces, highlighting why AI in games often lacks the complexity seen in research. The video also delves into performance limitations, AI decision-making processes, and the use of older AI techniques, such as state machines and utility functions, and why machine learning isn’t always a viable solution in game development.

Takeaways

• 😀 AI in game development is complex because of the many factors that must be considered, including performance, state space, and the need for quick decisions.
• 😀 There is a significant difference between AI trained to play a game (like AlphaStar in StarCraft) and in-game AI programmed by developers.
• 😀 Game AI uses symbolic AI techniques (like state machines and finite state machines), which are different from the machine learning systems seen in research.
• 😀 AI in games must make decisions very quickly, often in milliseconds, to maintain fluid gameplay and avoid performance issues.
• 😀 Games require a balance between intelligent AI and game performance, meaning that AI often has to cheat or use less intelligent methods to ensure the experience remains fun.
• 😀 A major challenge in AI for games is managing the vast number of unique states that can exist in a game world, especially in complex games like StarCraft.
• 😀 Determinism and observability are challenges in AI programming, as game states may involve randomness or partial knowledge of the state, complicating decision-making.
• 😀 Game AI often uses techniques like Finite State Machines or Behaviour Trees to ensure fast, predictable decisions, but this reduces the potential for more complex decision-making.
• 😀 Machine learning is not always practical for in-game AI during development because games are constantly changing, requiring frequent retraining of AI systems.
• 😀 Machine learning can work well in games with tightly defined problem spaces, such as racing games, where AI can be trained more effectively.
• 😀 While the goal is to make AI that is intelligent, the priority is to ensure that it is engaging and fun for the player, rather than being excessively perfect or predictable.

Q & A

What is the core reason why AI in video games is often less complex than the AI seen in research examples like AlphaStar or OpenAI Five?

The key difference lies in the fact that AI in video games is created by developers to work within the game's environment, accommodating varying skill levels, game logic, and performance constraints. In contrast, AI like AlphaStar is a machine learning system trained externally, using vast resources and data to excel in specific tasks, without the same resource limitations.

What is the challenge that game developers face when implementing AI that researchers don’t have to worry about?

Game developers face the challenge of integrating AI with the core game logic, rendering, input management, and online connectivity, while ensuring the AI operates efficiently within the performance constraints of the platform. This is distinct from researchers who focus solely on training AI to perform specific tasks.

Why does symbolic AI often get used in video games, despite machine learning being popular in AI research?

Symbolic AI is often used because it operates in real-time and can be integrated into the game engine without requiring external processing. Machine learning systems, however, require heavy computation and often need retraining if the game changes, which is not ideal for game development timelines.

What is a 'state' in the context of AI for video games, and why is it crucial for AI to manage states effectively?

A state refers to a snapshot of the game at a particular moment, which can include variables like enemy positions or player health. Effective state management is crucial because it allows the AI to make decisions based on relevant data and determine the best actions to take in the context of the game world.

How does the 'state space' affect the complexity of AI decision-making in video games?

The state space refers to all the possible configurations of the game world, and as the number of states increases, the complexity of finding optimal actions grows exponentially. In modern games, the state space can become vast, making it challenging for AI to calculate the best moves efficiently.

What does 'determinism' mean in game AI, and how does it influence AI behavior?

Determinism in game AI refers to whether the outcomes of actions can be predicted based on the current state. Non-deterministic games, where random factors affect the outcome, introduce complexity for AI, as it can't always predict the results of its actions or those of other AI characters.

Why is it challenging to maintain the performance of AI in modern games, especially as the game environment grows more complex?

Maintaining AI performance in modern games is challenging because the complexity of the game world increases, and AI must make decisions in real-time without causing lag. AI decisions need to be made quickly to avoid interrupting gameplay, while also considering performance budgets and system limitations.

What are Finite State Machines (FSMs) and Behavior Trees, and how do they help solve AI decision-making problems in games?

Finite State Machines (FSMs) and Behavior Trees help streamline AI decision-making by simplifying the logic for possible actions. FSMs ensure that actions are constrained to specific states, while Behavior Trees create a hierarchical structure for decisions. Both approaches balance efficient decision-making with game design needs, though they may limit the AI's ability to make complex, intelligent decisions.

How does Goal-Oriented Action Planning (GOAP) improve AI decision-making in video games?

Goal-Oriented Action Planning (GOAP) improves AI decision-making by allowing the system to search for optimal sequences of actions to achieve a specific goal. Unlike FSMs, GOAP allows for more flexible and dynamic decision-making, enabling AI to adapt to changes in the game world, although it is still limited by the complexity of the available actions.

Why is machine learning (ML) not commonly used for general AI in video games, despite its success in other fields?

Machine learning is not commonly used for general AI in video games because it requires significant training time, external resources, and ongoing adjustments. If the game is updated or patched, the trained AI system may become less effective and require retraining. Additionally, ML AI is more suited to controlled environments like racing games, rather than the dynamic and complex worlds of many video games.