I’m also interested in repurposing machine learning algorithms used for finding plausible hypotheses about data distributions into algorithms for finding action policies with high expected utility.
While I’m not in a position to hire you, I think that this is an extremely important problem. In the case where the utility function is known, I think there is lots of low-hanging fruit that will lead to progress in more “physical” application areas of machine learning like computer vision and robotics. In the case where the utility function is unknown, I think the problem is harder (at the level say of a PhD thesis), but would be a crucial step towards making progress on FAI.
If you’re interested in talking to me about either of these then I’d be happy to, assuming you have enough of a statistical background for me to get my thoughts across without too much of an explanatory burden. Assuming you haven’t already decided on a specific set of algorithms, I have some ideas here that I don’t currently have time to pursue myself that I think could lead to a publication in a good machine learning journal if done well.
I’ve moved in the direction of predicting AGI soonish (5-20 years)
This timeline is much sooner than I would predict. Could you perhaps point me to a few sources that you think would cause me to update my estimate towards yours?
There was a specific set of algorithms that got me thinking about this topic, but now that I’m thinking about the topic I’d like to look at more stuff. I would proceed by identifying spaces of policies within a domain, and then looking for learning algorithms that deal with those sorts of spaces. For sequential decision-making problems in simple settings, dynamic bayesian networks can be used both as models of an agent’s environment and as action policies.
I’d be interested in talking. You can e-mail me at peter@spaceandgames.com.
While I’m not in a position to hire you, I think that this is an extremely important problem. In the case where the utility function is known, I think there is lots of low-hanging fruit that will lead to progress in more “physical” application areas of machine learning like computer vision and robotics. In the case where the utility function is unknown, I think the problem is harder (at the level say of a PhD thesis), but would be a crucial step towards making progress on FAI.
If you’re interested in talking to me about either of these then I’d be happy to, assuming you have enough of a statistical background for me to get my thoughts across without too much of an explanatory burden. Assuming you haven’t already decided on a specific set of algorithms, I have some ideas here that I don’t currently have time to pursue myself that I think could lead to a publication in a good machine learning journal if done well.
This timeline is much sooner than I would predict. Could you perhaps point me to a few sources that you think would cause me to update my estimate towards yours?
There was a specific set of algorithms that got me thinking about this topic, but now that I’m thinking about the topic I’d like to look at more stuff. I would proceed by identifying spaces of policies within a domain, and then looking for learning algorithms that deal with those sorts of spaces. For sequential decision-making problems in simple settings, dynamic bayesian networks can be used both as models of an agent’s environment and as action policies.
I’d be interested in talking. You can e-mail me at peter@spaceandgames.com.