ricklentz · November 13, 2017 03:32
diff --git a/punt_external_algorithmist.txt b/punt_external_algorithmist.txt
 As a software developer, I struggled with the idea of external algorithmists. The idea that people will continue to build algorithms and that algorithms will be in use long enough before improving that people will review them seems odd. I guess people may continue to build them, but I will take a bet that the job I do will not be available for my kids.

 If you think this is all theoretical, I urge you to check out Planning Domain Definition Language (PDDL). It is a way to convert any deterministic problem into a search problem. After gettings hands on experience with PDDL, I realized that PDDL could be used to reproduce many of the decisions I make every day.

 Improvements in machine performance, like those Google's DeepMind published in Nature, have demonstrated that by using the same patterns, we can train machines to expert level performance across different problem domains. Specifically, reinforcement learning can tune performance iteratively based on new experience, better and faster than we can.

 Here is my private copy if interested in the DeepMind specifics: https://github.com/ricklentz/file_mirror/blob/master/nature24270.pdf
	As a software developer, I struggled with the idea of external algorithmists. The idea that people will continue to build algorithms and that algorithms will be in use long enough before improving that people will review them seems odd. I guess people may continue to build them, but I will take a bet that the job I do will not be available for my kids.

	If you think this is all theoretical, I urge you to check out Planning Domain Definition Language (PDDL). It is a way to convert any deterministic problem into a search problem. After gettings hands on experience with PDDL, I realized that PDDL could be used to reproduce many of the decisions I make every day.

	Improvements in machine performance, like those Google's DeepMind published in Nature, have demonstrated that by using the same patterns, we can train machines to expert level performance across different problem domains. Specifically, reinforcement learning can tune performance iteratively based on new experience, better and faster than we can.

	Here is my private copy if interested in the DeepMind specifics: https://github.com/ricklentz/file_mirror/blob/master/nature24270.pdf