What AI Training Lacks Now – A Curious Reflection

This article is motivated by Open AI recent demonstration in the Dota 2 scene. They aim to create a complicated environment to experiment AI training on. The AI became so good that it beats the 1st prize winner team of a Dota 2 biggest championship.

Now, I want to discuss something that hasn’t been touched on. When the AI plays with human, it plays a lot worse and often retreat to the jungle instead of its usual aggressive, fighting-centric, and objective-taking style. Why does AI have this weakness? And what is this weakness, exactly?

Then it clicked for me: Problem with AI training is it was trained to deliver the best result. Thus, only the fittest variations of the software survive, then get trained more to be even fitter, until it ultimately reaches perfection. When we incorporate AI with humans, it doesn’t work so well. Why? Because humans make errors. The situation that the AI finds itself in gradually becomes less and less than perfect. As more and more errors are made, the AI starts to tread on uncharted territory, and doesn’t know what to do.

This is similar to the young competitive chess scene Josh Waitzkin described in his The Art of Learning book. As a young chess competitor, he trains to think critically about the whole game while his opponents often memorize opening positions and traps that help them win quickly. As soon as mistakes were made and the game turns into an unfamiliar battlefield, it gets a lot harder for his opponents to play.

So regarding adaptive capability, human might has a slight edge over AI, at least until this point.

This brings up another question: Does it even matter that human has the adaptive capability edge? If human collaboration brings trouble for the AI, and the AI always complete their tasks perfectly anyway, should human collaborate with AI? Adaptability would become irrelevant, then.

Would there be any benefit to training the AI with random errors occurring as fail-safes for unlikely but potentially catastrophic events? Or would we rely on human to intervene in those times of need?

1 thought on “What AI Training Lacks Now – A Curious Reflection

  1. One of the big problems with training AI with reinforcement learning is that it teaches them how to achieve a goal, without ever teaching them situational awareness. It’s the why behind their actions that’s lacking. Humans make choices with long term goals in mind, though we are inherently flawed and often make poor choices, we can see the whole picture. Where as the AI can only assess the current game state with limited foresight. I don’t think training the AI with random errors is the right answer, but somehow training the AI to think strategically so they understand not only what is the best choice for right now, but what is the best choice to benefit 20 minutes from now.


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s