argos3 VS HungaBunga

Do check out these options (personally I’ve used these) 1. Buzz: https://the.swarming.buzz 2. Argos: https://github.com/ilpincy/argos3

You can do AI by relying on a library that is effectively a black box and yet make the result useful to other users. Recently I made a web interface for Argos3 (swarm robotics framework https://www.argos-sim.info ) and I do not know how it works. I understand in principle (write the code for a robot, define how many such robots run, how many ticks, start the simulation, study the output) but that is it, a very high level (some would say superficial) level. Yet, by providing a web interface (using Docker and Observable) I facilitated the research process. Did I do any AI? To somebody familiar with Argos3, no, to pretty much anyone else, yes.

Simulation is a strong choice too. ARGoS, Actin, Webots, Gazebo and V-Rep are some examples. A simulation will give you the opportunity to use some robots in different environments (with a lot of configurations). Gazebo has a big community and documentation, V-Rep is the more powerful one (and my pick).

How? Encapsulating the "core" part in an HTTP API. For example I don't write C++ and I don't even know much about swarm robotics. Yet to help a group of researchers I wrote a short NodeJS server for Argos3 https://github.com/ilpincy/argos3 that allowed few functions, in particular starting a simulation and get the results back. I did so naively at first and then in order to get faster feedback tinkered with server-sent events. This allowed me to connect an interface like ObservableHQ in JavaScript or Jupyter Notebooks in Python (but can also be JS, Julia, etc) by providing that API and results via HTTPS.

You can more easily understand errors. Not only why it happens but also what happens with them. Depending on your use case, you might save a lot of time and cost by selecting the correct model(s) in advance compared to using the HungaBunga classifier. Additionally, you might save a lot of time once/if the model doesnt work anymore. Here is an basic example that I have seen in the real world:

Sure, you can do your .fit(), get the model output and be happy. There are many cases where this works. But once you have to do specific models, change some things like the a custom error function to fit your use case better or simply understand why your model is doing things wrong, you absolutely need a good understanding of the maths behind it. You dont need to know all formulas by heart. But being able to understand them and having a good intuition separates the wheat from the chaff

But many dont. For many it is simply "use all of scikit-learn and select best". I agree that many things like the inner workings of NN architectures are probably not needed. But at the end, you are applying statistics. Knowing some basic statistics and math should be the norm.

There's a library that will go that for you: https://github.com/ypeleg/HungaBunga