Adding new variables dynamically

[rnbguy]

Currently, when initializing Reluplex, we need to pass the number of variables we are gonna use. The point is, it uses dynamic array to initialize the variables which can not be resized.

ReluplexCav2017/reluplex/Reluplex.h

Lines 189 to 195 in e2e48b2

	_upperBounds = new VariableBound[_numVariables];
	_lowerBounds = new VariableBound[_numVariables];
	_preprocessedUpperBounds = new VariableBound[_numVariables];
	_preprocessedLowerBounds = new VariableBound[_numVariables];
	_assignment = new double[_numVariables];
	_preprocessedAssignment = new double[_numVariables];

Does std::vector really affect Reluplex's performance?

[guykatzz]

Hi Ranadeep, I'm afraid I didn't understand the question. Currently the number of variables is indeed fixed during initialization. You could extend the code to support dynamic addition of new variables. Regards, Guy

…

On Thu, 21 Mar 2019 at 15:31, Ranadeep Biswas ***@***.***> wrote: Currently, when initializing Reluplex, we need to pass the number of variables we are gonna use. The point is, it uses dynamic array to initialize the variables which can not be resized. https://github.com/guykatzz/ReluplexCav2017/blob/e2e48b2980f76bdff74787427f43fff3b26a45a4/reluplex/Reluplex.h#L189-L195 Does std::vector really affect Reluplex's performance? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#10>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AOU33dtGYbmGyi9tkph55vnP4Diqkolaks5vY4mYgaJpZM4cBb0J> .

[rnbguy]

I am sorry if I was not clear. This is not particularly an issue - more like a feature request. I want to add new variables as I go. I am willing to extend it with std::vector. But before doing that, I want to know,

1. If this is possible just replacing these arrays with vectors and resizing them when I need new variables.
2. Whether using vectors will cripple Reluplex's performance.

[guykatzz]

It is generally possible, and resizing the arrays should not cripple performance any more than if those variables had been there to begin with (in general, additional variables lead to slower performance). However, correctly adding more variables on the fly would require a certain understanding of how Reluplex works, as you would need to adjust a bunch of data structures correctly. In particular, it will be important whether you were adding new equations or just new variables, whether the new variables participate in ReLU constraints, and whether there is already stored information about the execution that would need to be updated in addition to the current state.

You are welcome to email me directly and explain your needs, and maybe we can find a solution.