League of Legends. Cause they can't replicate the trash talk.
But what if the talk is a core part of the game? What if players win by talking well? Catan is the perfect board game to illustrate this. In each round, players negotiate with others to get the resources they need:
- A one-for-one trade: I'll give you one sheep if you give me one wood.
- A future trade: If you let me build on that tile, I will give you the next two wood cards I earn
- A sanction trade: I will only trade with you if you don't buy dev cards for the next ten turns.
- An agreement against trading: Hey, player B, let's not trade this round because it's better than trying to out-bid each other.
- Blackmail: If you don't make this trade with me, I'll steal your card.
How can computers ever solve a game like this? It needs to learn:
- How to understand the sentiment of others
- How to convince others to trust them
This seems impossible because a computer has to learn how humans think and feel, biases and all. They must learn how to prevent emotional retaliation because humans may not be playing to win.
Modern reinforcement learning algorithms achieve omnipotence by playing against themselves. So how can computers learn to play against humans if they only talk with computer players? These bots may even develop a new language to play Catan with each other.
Is Catan solved if the bot can only play against computers? I don't think so. I consider games as solved when a computer decisively beats world champions. Of course, we can train the computer to imitate human behaviour via supervised learning, but they will never achieve superhuman performance.
Despite all these challenges, I full-heartedly believe that an algorithm can solve these kinds of high human interaction games. There feels like a Turing award concealed behind this problem, and it's only a matter of time before it's won. I hope you have a lovely week!
- Curtis
P.S. I think this is a significant problem to work on because if we can create such an algorithm, the computer will be able to understand human psychology and rank the effectiveness of different ways of saying the same idea. A boring way to use this model is to ask it to generate supereffective email subject lines, so we no longer have to A/B test emails. A mesmerizing application would be to understand more about human psychology. For example, which phrases would increase the chance for nuclear deterrence to fail?
#94: Solving Games as a Service | #96: Homeless People Have Phones