Discussion Forum > AI Experiment in Minecraft - What Can We Learn?
I'm not sure if this is completely relevant, but this article, by trying to use AI to complete tasks usually done by humans in Minecraft, seems to have broken down tasks into components in an interesting way. It also argues that in a combination of "copying" (of what I see as skills) and explicit knowledge tends to yield better outcomes than either alone (my poor paraphrase). It also talked about reward functions, which I didn't quite understand. I wonder if there are lessons here for task/life management, where we're trying to solve "hierarchical tasks" in the real world? (Recognizing the limits of computer-human comparisons, as we've discussed before)
It’s common in AI development to rank a performance by the AI on a scale where higher numbers are better, and the AI is programmed to try to get higher values for this. The rank value is called reward, and the formula that computes the reward is the reward function.
The researchers in this project determined that the bot who learns both by imitation and by trying new things does best. Not surprisingly, as this can also be said of humans. But machines are not as capable of learning as we are and these things need to be proven and made precise.
https://arxiv.org/pdf/2112.03482.pdf