TLDR Tech / Practical
New metric 'agent plasticity' tracks AI learning efficiency
Researchers have developed a new metric called "agent plasticity" to gauge how effectively artificial intelligence agents convert experiences into lasting improvements. This is done by comparing held-out scores against the costs of learning. Evaluations across games like chess, Go, and Hex, as well as NetHack, revealed that the AI models achieving the highest final scores were not necessarily the most efficient learners. The study also indicated that even strong learning agents encounter difficulties in creating and applying knowledge that can be used in different contexts.
Read the full story