[HN Gopher] Evolving Reinforcement Learning Algorithms
___________________________________________________________________
Evolving Reinforcement Learning Algorithms
Author : graderjs
Score : 97 points
Date : 2021-04-25 06:11 UTC (16 hours ago)
(HTM) web link (ai.googleblog.com)
(TXT) w3m dump (ai.googleblog.com)
| neatze wrote:
| I fail to understand how such work should not mention neuro
| evolutionary argument typologies (NEAT).
|
| Does there's work substantially different then NEAT ?
|
| How does there's work compares in performance to NEAT ?
| nuclearnice1 wrote:
| Great questions!
|
| The blog points to their paper. Kenneth Stanley gets a mention
| in the first paragraph on related work.
|
| As far as difference, there are many. A major one is NEAT uses
| a direct encoding of the connections whereas this paper
| introduces this language describing the RL algorithm. Evolution
| is then conducted on trees where nodes are operators in that
| language.
|
| Performance comparison is complicated. A major issue is this
| paper finds graphs for doing RL. So it re-discovers TD or
| evolves an improved version of DQN. Whereas NEAT seems to
| evolve networks suited to specific tasks like pole balancing.
___________________________________________________________________
(page generated 2021-04-25 23:02 UTC)