Episode 24: Jack Parker-Holder, DeepMind, on open-endedness, evolving agents and environments, online adaptation, and offline learning
Generally Intelligent
00:00
Data Efficiency in Simulation Is Not That Useful
In most RL like benchmarks or training environments, the data is a lot smaller. We should be saying how good an agent can we get given as much computers we should find in simulation. But what we really know is being data efficient at test time which could be in the real world. So I don't care if it takes billions of samples to then be able to adapt quickly online when you do get that information. The key thing is I just don't adapt straight away.
Transcript
Play full episode
