Nobody is saying they shouldn't. They shouldn't talk about their sex lives to their students though, just like you shouldn't talk about your sex life to your coworkers.
Very nice write-up but most importantly very engaging visualizations which is key to draw in the reader and keep them hooked. Can see how it will be useful for someone who is coming in with nothing more than a few medium blog posts on "AI for journeyman" and no prior experience or education that one would expect to gain from a MOOC like Andrew Ng.
Funny and to the point. That’s kind of what I saw every time before. Now I’m learning with Fastai, and feel like finally someone made it easy for the rest of us.
The predictive analytics of these systems are strikingly close to reality. In fact, the AI is ahead of fantasy because the technology is capable of training itself through machine learning algorithms by using the data sets and past experiences.
Here I another article about: https://mobinspire.com/blogs/introduction-to-categories-of-m...
Will be glad to hear your thoughts on this.
This is a pretty good overview, though rather hand-wavy (eg the overview of RL here gives a VERY course idea of how this stuff works in practice). And the characterization of existing similar overviews as either overly technical or overly flimsy/fishy is a bit much ; this itself links to some things like Machine Learning for Humans (https://medium.com/machine-learning-for-humans/supervised-le... , or see eg http://brohrer.github.io/blog.html#how_machine_learning) , and in general as acknowledged by 'Yes, again' this is well trodden territory. Still, nice pictures ; could make for a great video , perhaps.
I guess you could say that there is an objective function to maximize and instead of thinking about generations you could consider the same individuals over and over again... That's my intuition of the similarities they could consider, but I agree with you that GAs are not RL.
I do not think the divisions so sharp. Many formulations in RL can be seen as implicitly defining a particular (set of) differential equations. One can do something similar for certain simple GAs and write them as equations of an evolutionary game where stable strategies are good solutions to the objective. The RL diff eqs also fall in this class.
Another view point is to note that John Holland, inventor of GAs, was more interested in the application of GAs in classifier systems than as objects of study in and of themselves. His work on the bucket brigade algorithm, a type of TD-learning, for credit assignment in complex reinforcement learning scenarios was first-rate and sadly still under-attended. In that setting, GAs were a search operator, focused on exploration while bucket brigade was for credit assignment. While GAs can be shown to be capable of doing adaptation as well as exploration, they really were meant to be part of a bigger whole by their inventor.
In fact, Deepmind's AlphaStar for Starcraft problem formulation can be seen as fitting into the learning classifier system framework where learners are themselves quite powerful neural networks. See: https://deepblue.lib.umich.edu/bitstream/handle/2027.42/2777...
Yeah, I agree and see where they're coming from, just wasn't sure if both RLs and GAs had been generalized into the same framework. Generalized to diff eqs is interesting.
21 comments
[ 5.3 ms ] story [ 66.7 ms ] threadFunny and to the point. That’s kind of what I saw every time before. Now I’m learning with Fastai, and feel like finally someone made it easy for the rest of us.
> thick academic trilogies filled with theorems ... or fishy fairytales about artificial intelligence, data-science magic, and jobs of the future.
> Genetic algorithms are considered as part of reinforcement learning
Uh, by who? Sure, there are similarities, but GAs != RL.
Another view point is to note that John Holland, inventor of GAs, was more interested in the application of GAs in classifier systems than as objects of study in and of themselves. His work on the bucket brigade algorithm, a type of TD-learning, for credit assignment in complex reinforcement learning scenarios was first-rate and sadly still under-attended. In that setting, GAs were a search operator, focused on exploration while bucket brigade was for credit assignment. While GAs can be shown to be capable of doing adaptation as well as exploration, they really were meant to be part of a bigger whole by their inventor.
In fact, Deepmind's AlphaStar for Starcraft problem formulation can be seen as fitting into the learning classifier system framework where learners are themselves quite powerful neural networks. See: https://deepblue.lib.umich.edu/bitstream/handle/2027.42/2777...
Yeah, I agree and see where they're coming from, just wasn't sure if both RLs and GAs had been generalized into the same framework. Generalized to diff eqs is interesting.