Researchers build AI that learns strategies by watching experts

In short: Researchers are training AI “agents” to figure out what makes experts successful by watching their actions, not by being told the rules.

What's going on

Some scientists are studying how to make AI agents learn by observation. An AI agent is a program that can take actions to reach a goal, like following steps in a task. The idea is to watch what skilled people do and then guess the hidden plan behind those choices.

One major approach is called inverse reinforcement learning, or IRL. In simple terms, it is like watching a chess player’s moves and trying to infer what they value most, such as protecting pieces or controlling the center. Instead of copying the exact moves, the AI tries to learn the “why” behind them, which researchers often describe as learning the person’s reward function (the internal score the system is trying to maximize).

Another related direction trains systems to infer goals and beliefs from behavior. That can mean predicting what someone will do in a new situation after seeing how they acted before. Researchers are also beginning to study AI behavior itself, treating AI systems more like subjects in an experiment to see how they change under different incentives.

What to watch

A key question is whether these observation-based methods can help with real-world work where people do not fully know how they succeed, or cannot clearly explain it step by step. It also raises a safety issue, since an AI that learns from watching could pick up habits or priorities that people did not intend to teach.

Source: NYTimes