Machine Learning For Games

Developing machine learning (ML) game mechanics

In 1997, history was made when IBM's Deep Blue defeats Chess Grandmaster Garry Kasparov in a six game series of chess. In 2016, history was made again when Google's AlphaGo defeats Lee Sedol in the  simple, yet hard to master game of Go. Only three years later, Open AI's Five beat the best Dota 2 players, shortly followed by AlphaStar beating the top Starcraft 2 players on the same year.


The progress made in the field of artificial intelligence is absolutely astounding, the message however, can be quite frightening:

AI Defeats Humans...

AI Defeats Humans...

AI Defeats Humans...

This doesn't sound like it will end well...

Maybe we should train AI to play relaxing games like Animal Crossing? Maybe we should train AI to play relaxing games like Animal Crossing?

If we can use games to make machine learning better... Why can't we use machine learning to make games better?

We could use neural networks to approximate  answers quickly, replacing the need for laborious and expensive computations. Neural Networks can also be used to create dynamic, immersive animations that adjust accordingly to the environment in vivid and believable ways.

Imagine having characters that can adjust their actions on the fly, creating emergent behavior. Characters that learn can enhance many existing game genres and open the door to new experiences.

What if in Nintendogs or on a Tamagotchi handheld, instead of having a pre-programmed collection of tricks, the pet could truly learn new tricks?  Or how about a sports game, where you coach your team to success by teaching them how to play?

The possibilities are practically limitless. Though the road leading to machine learning with games, is a long one indeed.

When your friend's internet dropped out, could we swap in an AI replica of them and continue the game? When your friend's internet dropped out, could we swap in an AI replica of them and continue the game?

The Road Ahead

So what are the unique challenges in developing machine learning for games?

Here are some we're facing at Strife AI LLC:

  • Making ML training fun for players: How do we incorporate this technology into fun experiences?
  • Making ML explainable to players: If the training isn't working, we want players to understand what they are doing wrong.
  • Doing more with less data: If most ML algorithms require astronomical amounts of data, how can a character learn anything interesting from one player?
  • Doing more with less powerful hardware: We can't expect players to own a data center, and we can't provide one for each player either.

We have already begun exploring solutions to many of these challenges in our game prototypes, and we are convinced that unique games are feasible with adjustments to today's ML techniques. We will use this blog to explain how we are confronting these issues and more.

If you are a game developer interested in using ML in your own games, please contact me at brendan@strife.ai


More posts