Reinforcement learning is a method for learning incrementally using interactions with the learning environment. It is an approximate and incrementally improving solution to an optimal control problem (sometime called approximate dynamic programming).

Casual inference is totally different. Causal inference is trying to learn causality from data (what was the cause and what was the effect).

Thank you for presenting this important topic....

I think the fuzzy logic techniques can be used to solve all problems related to causal inference.....

good luck

Peter Jardine ,

Lets start with the caveat that Pearl's interpretation of causality is not widely accepted. For example , for me, it remains just probabilities until a explanatory mechanism is achieved that you get causality.

That being said, I think that your difficulty lies in the interventional process described in:

https://www.inference.vc/untitled/

The intervention as described in the article refers to the actual intervention of the data and not on the ML algorithm's computation of its observations.

"What causal inference offers that reinforcement does not?"

The difference lies in that in causality you are trying to model the underlying behaviour of the system under observation. Under reinforcement learning you are trying to model a belief based on what is being observed. Beliefs do not necessarily model the underlying behaviour ( missed observations, etc.), though they might sometimes converge to the same result.

Dear Arturo Geigel . Your question and puzzlement about the terms causal-reference versus deep/reinforced-learning is very much in place! as long as using the Statistics paradigm, the terms are "the same lady with a different dress". See No Matter What You Call It, It’s all the Same Thing www.linkedin.com/pulse/matter-what-you-c...pzdHm4ySV0jjZg%3D%3D

My solution is to use the new data science instead, to unlock the chicken-and-the-egg cycle where unchecked assumptions instruct the learning.

Note that human learning has the same issue of poor results, but machines are doing it faster and "better", so it surfaces only now.

In my view, the current scientific model is missing a piece that should guide hypotheses making. This is the center of my project https://www.researchgate.net/project/Philosophy-of-Data-Science-review-for-big-data-analytics - see the log.

This is something I am currently looking at, and my intuitions match yours.

Indeed, looking only at data we can identify V-structures, and this is all we can do with regard to inferring causality. Indeed, RL algorithms are experimenting on the environment and storing causal information, at least implicitly. I am just hesitant to say that this happens in models fulfilling the Markov property, it needs further analysis.

Tomasz Dryjański I agree that, intuitively, RL seems like it should be storing causal information because it is producing rewards through interactions with the environment. That said, as Arturo Geigel mentions, RL is really just testing a belief to find some optimal point in the search space. At least in the traditional framework, RL is not actually modeling the underlying cause and effect. Thanks, team - your responses have helped point me in a better direction.