Beyond Alpha Zero, where is neural networks reinforcement learning now?
Sort:
No
this is reinforcement learning, the ai isn’t actually learning, it just tries different concepts and gets rewarded if it is (In the hiders case) not seen.
A neural network is much more complex
Jun 17, 2020
0
#5
redghost101 wrote:
A neural network is much more complex
If you go through the paper the video is on, you'll see that they used deep reinforcement learning which basically makes use of both deep learning and reinforcement learning principles. So they actually did use a neural net.
Jun 17, 2020
0
#6
redghost101 wrote:
this is reinforcement learning, the ai isn’t actually learning, it just tries different concepts and gets rewarded if it is (In the hiders case) not seen.
But that is learning though.
Jun 17, 2020
0
#7
The more advanced opponents it plays, the stronger it gets.
Therefore, AlphaZero can get uhhh.... unpredictable as the neural networks are dependent on the opponents it plays. If it plays a 500 rated player over and over again it learns from the match after playing. The computers logic - I need to just beat that particular human, not the world. AlphaZero
cant just play with weak opponents, else it becomes weak. The more it plays with itself..... The better it gets at drawing itself.
Me no professional, but as far as I know. Reinforcement learning is when they try every move possible until they get the highest reward, then continue onto the next stage. Neural networks find how to get the next move, making them more efficient than reinforcement learning
The thing is, neural networks don’t learn from playing 500s, they learn by playing themselves. One NN as black, one as white. Once the training phase finishes, they then challenge much higher rated people or a.i
It can take from 10 hours to 10 days to train, but once it does. It’s almost unbeatable
Good point!
Ta very much
@MarcoDiazz it is not learning, this is an algorithm. A neural network is not
Jun 17, 2020
0
#15
redghost101 wrote:
@MarcoDiazz it is not learning, this is an algorithm. A neural network is not
?
Flip it
Jun 17, 2020
0
#17
redghost101 wrote:
It can take from 10 hours to 10 days to train, but once it does. It’s almost unbeatable
According to common logic, playing by yourself does improve your ability. i.e if you play with others lower than your rank, YOUR rank depletes.
It plays itself, each time trying a new strategy. BOth sides try it, letting the. NN see it’s weaknesses and improve ovqer time
Jun 17, 2020
0
#19
redghost101 wrote:
It plays itself, each time trying a new strategy. BOth sides try it, letting the. NN see it’s weaknesses and improve ovqer time
hmmmmmmmmmm....
When Alpha Zero scientific experience end in Chess Go and Shogi, many neural network training are going on many games like Starcraft, Dota and many other games.
This is one of achievement where neural networks can learn and evolve exactly the same like human.
https://youtu.be/kopoLzvh5jY