Scripts and Rules: The Future Is Not Fully Automatic
September 24, 2018
I wanted to capture an item of information which may be lost in the flood of marketing craziness. The subject is smart software, autonomous systems, and why humans may have to push buttons and turn knobs.
The write up is “Deep Learning Is Inferior to Scripting for Teaching Robots.” The main idea, as I understood the article, is that creating useful robots (hardware and software) may benefit from old school methods.
The method is for one or more smart humans to create data sets and train robots. But the humans are not out of a job. The robots have to be retrained with more rules, updated rules, and fresh data sets.
The article points out:
They [A. Rupam Mahmood, Dmytro Korenkevych, Gautham Vasan, William Ma and James Bergstra] conclude that deep learning lags behind the traditional way of training robots with scripts ‘by a large margin in some tasks, where such solutions were well established or easy to script‘. However, RL was ‘more competitive’ in complex tasks (for instance, docking to a charging station). The report also shows that the RL algorithms need careful tuning to their variables before being useful for any task – but after tuning, those same variables were applicable across a range of tasks.
In short, autonomous methods are not as efficient as humans doing the rule work and creating scripts or training sets.
The point is an important one. Smart software is not the cost and accuracy silver bullet that marketers have described as a reality.
I am not disputing that for certain specific operations on bounded data smart software can be magical.
But bootstrapping intelligence and learning with zeros and ones has not yet docked at the port, unloaded, and moved the goods to the user.
Stephen E Arnold, September 24, 2018
Comments
-
- Subscribe to Beyond Search
Feature archive
News archive
-
