And they said video games don't teach you anything. BACKGROUND:
This game immediately caught my eye given its subject matter being something I actually learned about in college. Some of my friends have actually played this as well. My turn. PLOT: Our protagonist is an aspiring programmer going through the all-too-familiar struggle of dealing with bug-riddled code. One particularly frustrating moment causes them to walk out, at which point the pet cat takes over and... solves the problem? Bewildered (because who wouldn't be), our main character resolves to create a cat-to-human translator to find some answers. It's a silly premise that gives the main character a reason to pursue the field of machine learning, and the rest of the game carries that air of silliness via the email requests you receive from clients who have need of your slowly improving coding skills. Those emails make cheeky references to pop culture or simply keep with the absurdity of the story by having barely coherent emails clearly written by cats, though the requests themselves fall in line with some of the real-life applications of machine learning (like image processing software). That said, I can't really say that I was particularly invested in the end goal of figuring out what the cat is saying or found most of the jokes to be to my taste. The game takes time to give a quick overview of the history of machine learning between puzzles, which is nice and informative, but as somebody who actually took a class that covers the subject, I can say that it's far from comprehensive. PRESENTATION: The visuals take on this simple, cartoony style with flat colors and little to no fine detail. It's cute, but I can't say that it really captured my imagination or anything like that. The UI for the main gameplay is pretty clear, but I did have issues with the lack of space you can work with, since you have to zoom out into space to fit in several elements. The music is more or less the same throughout, and it's this simple xylophone (or something equivalent) melody that sets a chill vibe for the gameplay. I found it repetitive after a while, so I ended up switching out to my own music/podcasts. GAMEPLAY: While True: Learn is a puzzle game heavily inspired by the concept of machine learning. The idea is to take a data set (oversimplified to just be shapes, colors, or a combination of both), run it through "nodes" that represent various types of machine learning algorithms (which process the "data" differently), and transport the data to the appropriate receptacle (signifying that the algorithm has interpreted the data correctly). Pretty much all the puzzles involve simply setting up the right nodes and having them detect the right parameters such that the whole "program" sorts everything into the right receptacle. This is an oversimplified take on how machine learning applications are actually developed. At no point do you actually learn to create any of the machine learning/AI algorithms. The most you'll get is a summary of how they work, and even that is criminally abridged; I only know what the hell any of the models are because of prerequisite knowledge. I think of this game more as a visualization of how machine learning algorithms process data, which does have value in that it shows that such algorithms aren't perfect right out the gate. Unfortunately, you're not really going to learn how to put together specific machine learning models, you don't actually train any of your models to learn anything from your data set (which is a major component of machine learning), or even learn how to ensure your data set is collected properly to begin with. If you want to learn more about the machine learning concepts the game is invoking, you are directed by the game to click some outside articles or videos regarding the subject. This approach was incredibly bemusing to me, as I went into this hoping that the game would teach you about the concepts with its gameplay. That does not happen at any point, and I think that's a massively missed opportunity for a game that sells itself on the notion that it's a gateway to machine learning. Rather, the objective of the puzzles of this game is to have your setup of nodes deliver the data to their designated receptacle in a timely and accurate manner and making sure the quota of each receptacle is met to begin with (this is either quantity of data or maintaining a specific threshold of accuracy for a specific time). The longer I think about it, the weirder this approach is to me, since the game quickly becomes all about optimizing your setup as opposed to simply solving the task at hand. I don't know about anyone else, but I typically worry about my code accomplishing its objective first before I hit my head on the table over and over trying to figure out how to make it run better. This becomes an issue to me due to the existence of medal ratings for each puzzle, which incentivizes you to make sure your setup sorts the data within a time limit and encourages you to use only a specific number of nodes. It feels like you're actively punished for making the most out of your available tools (your medal rating lowers the more nodes you use) and by not immediately figuring out the most time-efficient setup (some of the puzzles feel straight up unsolvable because of time limits). I couldn't put my finger on what was frustrating me about chasing for gold medals until I sat down to write this, and what makes this worse is that the reward for getting many gold medals is practically meaningless (spoiler: it's just infinitesimal variations of story cutscenes). Now that I think about it, the game is more a collection of math-based puzzles than anything. Often, the key to figuring out the best setup of nodes involves carefully checking your data set (as the quantity of each object is specific), understanding the sorting behavior of the nodes themselves, and checking the conditions for the receptacles they'll be going into (like the data object quota and the percentage of "correct" data). You must then set up your nodes such that all those conditions are satisfactorily fulfilled. I didn't mind this being the conceit of the puzzles (math is fun to me, too), but it's certainly far removed from what I assumed was a programming game. In terms of time limit, each node has a specified running time, which represents how long the node "processes" a data object before it spits it out to the next node. It hardly seems relevant since there's no real way to manipulate a node's running time, but it can become an issue since nodes have a max data capacity, and if the node doesn't process data fast enough, the flow of data can get throttled, slowing everything down. Figuring out how to get around this was an exercise in frustration, and that's not even taking into account setting up nodes such that you get the gold medal time limit. The closest thing this game has to machine learning are the reinforced learning puzzles, where you try to teach a self-driving car how to detect obstacles and move accordingly. This is more in line with what I had in mind (visualizing how machine learning models learn by iteration), but the execution of it in this game left something to be desired. There was no clear indication as to how the car's AI was learning from my training, so the puzzle quickly devolved to clicking shit until I got what I wanted. Moreover, the main criterion for clearing the puzzle is ensuring that the car maintains a specific average speed. Now, I'm not an expert on self-driving cars, but I would assume the AI of such a car should prioritize not slamming into anything over maintaining a speed limit. You earn money when you clear each puzzle, which you can spend in the shop. Most of the items are useless cosmetics that only exist to change the main screen, but you can also buy upgrades to your character's setup to optimize their machine. These, in turn, improves things like data transfer speed and such, but as far as I can tell, these improvements are barely noticeable. Also, the existence of this mechanic implies that the moral of the story is to throw money at better hardware to programming problems, which is.... not exactly the lesson to take away. There are also "startup" puzzles where you have more leeway with how many nodes you can use to solve the problem. You can then push your finished setup and the game determines a profit based on how "customers" feel satisfied with the product (and patch things accordingly, if needed). It's a good way to earn some extra cash in-game, but again, there's little to spend money on. Frustratingly, speed is still the main metric of success (people get mad if your application is slow, and you'll soon find yourself operating at a loss), and the game doesn't tell you what an acceptable running time is (making the former problem extra infuriating). Also, every startup seems inevitably doomed to fail (too real?), and you have the option to cash out at any point before then. VERDICT: I really wanted to like While True: Learn, but I found the dissonance of what it's trying to be (a machine learning course in the form of a game) and what the game actually has you do (solve puzzles all about optimizing the flow of data as opposed to piecing together machine learning algorithms) too jarring to ignore. I don't have a problem with the puzzles conceptually, but I wish it leaned more towards being open-ended with solutions (which the game permits to an extent) instead of frustrating me with trying to figure out the one optimal setup. It also barely teaches you about its own subject matter, which is a shame because I feel like a game that makes machine learning accessible to people outside computer science would have value. I don't really know how a game that does what I hoped this game would do would even be structured, so I'm sadly left frustrated at how it's only tangentially connected to machine learning. If you're looking to get this game in the hopes that you'll come out of it knowing a thing or two about machine learning, I'm afraid the most this game has to offer is an abridged history lesson and an oversimplified visualization of how some machine learning models work. As a game itself, I feel like the puzzle gameplay is an interesting concept, but I think the way success is measured may lead to frustration for some players. Thus, I have some reservations about recommending this game to people. Choose your preferred platform: - end -
0 Comments
Your comment will be posted after it is approved.
Leave a Reply. |
Categories
All
Archives
June 2024
Derryck
|