AI planning algorithms in RustJune 14th, 2020 • No Comments
Artificial Intelligence (AI) is a hot topic, although it seams that the focus is on Neural Networks & Deep Learning on the software/algorithmic side, while GPUs/TPUs are the #1 topic for hardware. While this is fine, it feels sometimes like other parts of what had been part of the overall AI domain in the past fall short (See the book Artificial Intelligence: A Modern Approach). Just by looking at the table of contents you can see many more AI topics which are critical but do not get a lot of attention atm. For example planning & reasoning over the insight you have gained – by e.g. using neural networks – is essential to build “semi-cognitive” autonomous systems. Btw I believe planning (the what/how) is as important as figure out the scheduling (when/where) – for example AlphaGo Zero also uses a planning component (using Monte Carlo tree search) for figuring out the next move – next to a neural network.
To have an excuse to learn Rust and to get more into planning & reasoning algorithms I started developing this library (Warning: this is still a work in progress). The experience with Rust and its ecosystem have been great so far. The Ownership concept takes some time to get used to, and probably only makes total sense after a while. A lot of the other concepts can also be found in other languages, so that makes starting easy.
The following example shows the D* Lite algorithm in action. Assume you have a maze like the one shown in the following figure and the goal is to move a robot from the coordinates (5, 0) to (0, 5). To make things more complicated, we assume a path along the corners of the maze is more “expensive” – so at the start the path a robot could take looks like this:
In contrast to A*, D* Lite uses a backward search and enables re-planning – that means that after e.g. the robot has moved two steps (to coordinate (5, 2)), and an obstacle gets removed (coordinate (3, 4) no longer represents a wall) it can re-plan without the need to recalculate all values (saving compute time). This can be seen in the following diagram (both diagrams have been plotted through rustplotlib by using Matplotlib in the background, btw. – this shows how easy it is to integrate Python with Rust.):
Why Rust? Over the years I learned some programing languages, some probably gone extinct by now. This includes candidates such as Oberon, Turbo Pascal (anyone remember Turbo Vision?), Elan, and Basic. I used to code in Java & C/C++ for a while – today it is more Python and a little bit of Go. It has been nagging me for a while, that I wanted to learn a language that was somehow “revolutionary” and offered sth new. For example Java is a great language & has a powerful SDK; Python is easy to use; etc. But Rust finally offered something new – the concepts around its memory management (w/o the need for garbage collection) & the way that it enforces you to write clean code are compelling (Maybe we can finally battle cargo cult programming). I have looked into Erlang too btw, but the syntax is just a bit too quirky – however the OTP is a very strong concept and sometimes it feels like that Cloud Native Applications + Kubernetes want to be the future OTP 🙂 Microsoft btw also describes some of their motivations to use Rust here.
So Rust is perfect? Probably not (yet) – many tools/libraries in the ecosystem are still young (e.g. grcov works only with nightly toolchain), and do not offer the same range of capabilities as e.g. in the Python world (still looking for good machine learning library etc. for example). Getting to know Rust is a bit painful in the beginning – and one word of advice: do not always follow the tips from the compiler, which it spews out on errors – sit back, think about what the errors actually means and fix it accordingly.
Any Book suggestions? So for learning Rust I mostly went with the “official” Rust book and got an older hardcopy of the book (the last programming books I bought before this was on Erlang years ago btw.). I heard the O’Reilly book on Rust is great too btw.