AviationGuy wrote: ↑22 May 2019, 09:31
AviationGuy wrote: ↑21 May 2019, 14:57
Only data for the input boxes or could the AI database also provide information
for the RPA bot on where
the RPA bot needs to click on the screen based on what's on the screen?
Is it better when I phrase it like this?
...Since AI comes up with predictions, they are not always right. This isn't a problem really when this happens with a chatbot.
But what if... let's say, an RPA bot is responsible for ordering products based on the data it receives. This data will mostly be valid and useful, but in some occasions the amount of products that need to be ordered isn't present in the data and an AI needs to predict the amound based on previous data and passes this on to the RPA bot. Because, this amount is based on previous data, it will mostly be an acceptable amount but this won't always be the case. The wrong amount can be ordered because of the wrong prediction made by the AI. When you as a company work with high value products you probably don't want this and thus will rather not choose for an Intelligent RPA bot. Although, you can still use the Intelligent RPA bot but you'll need a human to check what the IRPA bot ordered. Is this then still preferable over an RPA bot which simply sends the data without an given order amount to a worker which then needs to order the product himself?
I hope you understand what I'm trying to say.
When is an IRPA bot really a benefit? And when do you still want to use a 'normal' RPA bot which let humans performs the tasks it isn't able to handle?
Maybe you might want to think of AI differently. When people are referring to AI, they are often referring to "narrow intelligence" or "weak AI". A lot of times, the fact that a program is automating something or can function by itself, many people that don't understand programming will often label it as "intelligent" or as AI. So we have to be careful with the term. The program can simply be following a set of rules or selecting options based on predetermined conditions, thus is automating a process.
A program can be excellent at "making decisions" based on predefined rules or statistical probabilities, within a specific scope, but not so in a general sense nor come up with creative solutions for a problem. If a task, situation, or parameter is outside the scope of what the program was made for or can/should handle, then a human has to intervene. When a program or robot can learn on it's own, in more generalized ways like humans do, then we might be talking about real AI. But where RPA stands now, we are usually referring to automation, more than AI.
Let's say you had an "intelligent program" that was designed to select pictures of certain fruits from any unknown source. It would determine if a picture contains a fruit by shape, colors, or patterns. Since we would be dealing with statistical probabilities and unknowns, this would fall under the umbrella of AI or Cognitive Automation. If on the other hand, the options and choices were set and known (selecting from a known set of pictures, colors, and shapes), this would usually be seen as Traditional RPA or just automation.
In the case of AI or Cognitive Automation, humans would still be involved in training and the setting of acceptable ranges of accuracy. If the program is certain at say 96% or greater, then it could mark the picture as the desired fruit. Below that percentage, it will mark as not desirable or not a fruit. This is the "intelligence" we are dealing with, where the program has to "guess" the right answer.
It can happen that the program gets some pictures wrong, because they are a new type. Something occurred that was outside the original design and capability of the program. A human would have to intervene to train/modify the program to recognize the new kinds of pictures.
If the program can sort through pictures of fruit 30X faster than a human, with say an accuracy rate of 99% or greater. Then that leaves just a small percentage of pictures that must be checked manually. And the human can retrain or modify the program to attempt to get the error rate even lower. How well they could raise the accuracy and reduce exceptions, would depend on many factors. Then for many people or companies, that might be a very acceptable situation. Less people or less hours are needed to sort through the pictures.
Also, a program has to go through testing to make sure it's working as intended. If the acceptable level of accuracy must be 99%, then they would not release or approve of the program until it showed that level of accuracy in testing. If the program is used by a company, the process would be designed so they wouldn't lose money or create a disaster. In those cases where the program was not sure, then it could be turned over to a human to check. Nevertheless, the amount of work or hours that humans had to do would be reduced, even if they had to handle the exceptions. And the program could be revised to reduce the number of exceptions, based on the knowledge and experience they have accumulated, but this can take time.
Furthermore AHK can even be utilized in the machine learning portion. Most RPA software cannot. Blue Prism for example must interact with Watson for AI. AA is pretty limited in its built in machine learning.
Echter, deze nog niet! Zal snel is langs gaan voor een