When I think "optimal language", I don't think "optimal combination of natural languages". This is partly because I've been experimenting with constructed languages (conlangs) since the age of eight and partly because my ethical and aesthetic desires lead me to different (or at least more elaborated) goals than "better communication between the peoples of the world as they exist today".
There are many conlangs for various purposes. Esperanto and Interlingua are useful for facilitating communication between people, acting as geography-blind APIs. Lojban is useful for communication between humans and computers, as well as fostering a strict logical mindset similar to the effects of working with programming languages, mathematics or symbolic logic. Klingon, Elvish and many other languages of various semantic/syntactic comprehensiveness are created for aesthetic and/or fictional purposes.
I've been working through a new conlang idea in my head the past few days. It builds off of Lojban in that its tenets are (1) universality, and (2) ease of I/O by both humans and machines. But it adds onto that (3) an easy-to-use, potentially unconscious system of real-time abstraction.
We do this to some extent with all human-human communication, when we inflect or affect or twitch or sigh or use other non-verbal linguistic cues to impart semantic value onto entire complex utterances. The way that a hyperlink imparts the contents of its destination onto the meaning of the information between its opening and closing tags.
This could be significantly sped up through use of more standardized multi-sensory linguistic components, either attached directly to individual people or attached as a metadata layer readable by nearby people through an AR technology of some sort (glasses are the obvious choice, but let's not limit ourselves).
For instance, a speaker could say the word "lu" multiple times in the same monologue. Those around the speaker would grok the intended meaning of "lu" each time due to whatever technological cues were broadcast along with it. The cues could be a string of words on a HUD or a color tinting the image of the speaker or a soft harmony in the listener's ear or a distinct pattern of buzzes from piezo elements in an ankle bracelet or even direct electrical stimulus causing a particular mental state.
I'm getting ahead of myself.
The point is that this is the sort of thing I imagine "language optimization" to be. And this example is just scratching the surface of potential optimizations, because any optimization that allows for general abstraction opens up all the doors of Turing completeness.
Esperanto and Interlingua are patches. They are certainly useful for some things[1], but we can't know what they're optimal for unless we're also experimenting with entirely new systems not based solely on natural languages. We need to be able to treat real-time language as scripting — anything less and we'll be limited by the language rather than by our cognition.
https://en.wikipedia.org/wiki/Esperanto