NN for solving the Rubik's Cube would be like killing a mosquito with an artillery bomb: not necessary at all.
-
-
Although not related to the OpenAI paper. Solving Rubik's Cube via Deep Learning shows that you can teach a system to learn how to solve the problem. That's a very different problem that programming the God algorithm.
1 reply 0 retweets 2 likes -
Replying to @IntuitMachine @dmonett and
Isn't that a trivial observation about sequences? More interesting is whether i) it is cost efficient, ii) the solution is generalisable, iii) whether we learn anything from the solution. You can dig a hole with sieve but is it the best tool?
1 reply 0 retweets 1 like -
Replying to @markburgess_osl @dmonett and
Training a system to implement an algorithm on its own is a different problem than formulating a solution by hand. Now if a DL can discover the God solution, then that's a massive breakthrough.
3 replies 0 retweets 0 likes -
Replying to @IntuitMachine @markburgess_osl and
Go a step back: when you already know the solution and the exact steps to get there (plus how to do it in a much simpler and straightforward way), you don't need
#DeepLearning to figure it out.1 reply 0 retweets 2 likes -
Replying to @dmonett @markburgess_osl and
You don't know how a cube is scrambled. That's not a given.
1 reply 0 retweets 0 likes -
Replying to @IntuitMachine @dmonett and
An automated algorithm to solve the Rubik's cube took humanity 7 years to discover: https://en.wikipedia.org/wiki/Optimal_solutions_for_Rubik%27s_Cube …
1 reply 0 retweets 0 likes -
Replying to @IntuitMachine @dmonett and
Solving Rubik cube through RL is already done before and so I believe
@openai goal wasn’t that. The focus was real-world manipulation task and sim2real.https://www.nature.com/articles/s42256-019-0070-z …2 replies 1 retweet 6 likes -
Replying to @sytelus @IntuitMachine and
ps deep rl + search ≠ deep rl. that paper to is a hybrid.
1 reply 0 retweets 2 likes -
Replying to @GaryMarcus @IntuitMachine and
Hopefully MCTS band-aid will go away some day. However I felt blog+paper were very clear on using Kociemba’s algorithm. Also pleasant trend of describing open issues, i.e., (1) 80% attempts fail on hard config (2) custom cube communicates state (3) 13K years cummulative training.
2 replies 0 retweets 1 like
any why characterize MCTS as a band-aid, diminishes its value? it's a great tool that gets used over and over; why the prejudice against anything not neural net? it's like someone owning a hammaer dissing a screwdriver as a "band-aid".
-
-
Replying to @GaryMarcus @IntuitMachine and
MCTS is powerful tool indeed. I think many of us are more obsessed towards purely learned approaches with hope that they are more adaptive/longer term robust over diverse problems. It may not be good obsession. MCTC/other tools might indeed turn out to be essential primitives.
1 reply 0 retweets 2 likes - 1 more reply
New conversation -
-
-
Replying to @GaryMarcus @sytelus and
The prejudice against anything NOT neural network is because the ultimate goal is that anything can be learned.
0 replies 0 retweets 0 likesThanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.