"Learn how machine learning really works!" says the game's blurb, and the puzzle descriptions sound right up this alley. The challenge is to send the inputs to the correct outputs using a minimum of blocks and within a time limit, so you may have to spend some time optimizing your approach.Īt the same time, though this game is all about machine learning. You do this by building a flowchart made of conditional blocks: a "Decision Tree: Color" can send one color to the top output and another color to the bottom output (the third color is sent randomly to either output), a "SIFT" sends one shape to the top and the other two shapes to the bottom, a "MARK-1 Perceptron" can divide the colors across its three outputs but requires training (and misidentifies some objects), and so forth. I spent seven hours getting to the end of this game, and I'm of mixed feelings about it.Īt its core, While True Learn is a puzzle game: given bins of input objects (red/green/blue circles/squares/triangles), move them to the output bins that want them (like, one bin might want red circles and blue squares). The puzzles stop being interesting very early, and those that attempt to introduce more complicated machine learning topics avoid the hard task of introducing it in a way that the user can learn about the topic via the puzzles. The RNN puzzles give you the idea of RNN, but don't allow you to connect the nodes in any interesting or meaningful way: each node has memory, but you are forbidden from doing anything with it other than feeding it back into the same node that produced it. The RL puzzles give you a lot of knobs to play with, but the puzzle doesn't require anything other than sending some inputs to the output stream. In particular, the reinforcement learning puzzles and RNN puzzles don't let you use the ideas to come up with clever solutions. Most of the machine learning concepts introduced in the game end up being irrelevant to the puzzles. This is annoying in combination with the prior point, when you are trying to debug which stream is getting too many nodes so that you can balance the output. In these cases, if you want to see what went wrong, you will have to wait for the test to time out until you can view the stats for each output stream. The test run does not always terminate when the test is unsolvable, for example if all inputs have been sent to output streams, but the current accuracy and / or volume in the streams is not adequate. Most of them fall under the "load balancing" category, which means that you spend most of your time figuring out how to allocate output symbols to output streams to match the volume dictated by the puzzle. After a short time, the puzzles become very uninteresting. The game gives useful links to those interested in learning more about machine learning. It's also easy to use solutions from past puzzles as nodes in your current solution. The visual editor for solving puzzles is easy to use. I'm sad to say that past the first hour or two, you'll have seen everything of value in the game. I played it to the end to try and see what it had to offer. Later levels add side-facing LIDAR so your car's AI has more information about when you pass another vehicle! Except this only confuses the poor AI and removing the side LIDAR solves everything. Each level adds more sliders which tune the way in which your AI evolves - but you just ignore them and click the evolve button. The solution to all of them is to just click the evolve button once and everything's done. They claim to use a genetic algorithm to evolve the AI. The worst offender is, sadly, the driving simulations levels. The game expects you to briefly hook these up to your system with an extra "gradient descent" node to "teach it", but once you do this, just copy that node to all the places you need it to avoid "retraining" them. Late game pieces are not fun to play with - because the game grades you on "processing time" you get faster pieces with small error chances. Except for that one level where it forces you to use a previous level's solution. Levels randomly restrict which nodes you can use so while they added the ability to reuse your solutions from previous levels, odds are that level used a node you're not allowed to use and you'll just have to do everything from scratch. I struggle to even call this a puzzle game when so many of the levels are putting a single piece in the middle of the board and connecting its ins-and-outs to the board.īuying better hardware means you'll run the same solution you had on an earlier level and upgrade from a silver star to a gold star - not because you made anything better, but just because "your computer is now faster".
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |