Just out of curiosity, how are you going to approach the "optimized curation bot"?
The reason I ask is because I have my own curation bot running on aicu, it's doing semi decent, but it has still lots of room for improvements.
Posted using Partiko Android
I already got one running on this account if you wanna check efficiency on Steemworld, which is pretty decent too (approx 125% in efficiency).
My idea is more to open it to delegations with return in powered-down rewards for delegators. On the bot itself there's of course always room for improvements, but with the current level of return, it would already be more than enough to attract delegators.
Interesting, your approach seems to work pretty good. I noticed that you are using pretty low upvote weights. How is your bot selecting posts ? Are you using user submitted posts or artificial intelligence techniques?
My bot is at 93% efficiency right now. But I never trained it to identify high value posts. Just high quality ones. So just from that an efficiency of 93% is pretty decent.
For that I'm using traditional machine learning and NLP techniques ( just a simple Latent semantic analysis combined with a random forest). I'll upgrade those techniques with state of the art language models etc soon, but for now that's my approach.
Posted using Partiko Android
Bot selects the vote weight that maximizes the curation return. But if you take only that, you'll always end up voting at 0.01%
The thing is that the bot tries to maintain a certain curation efficiency. Depending on its current efficiency (pending rewards) it will take more or less risks (vote a bit sooner, without trying to maximize those few more seconds, or vote a bit later at the risk of having a big vote before yours, then cancelling altogether the vote on that post as it's now under the efficiency threshold)
The higher the efficiency wanted, the lower the vote weight will be
But VP also comes in play, a high VP will make the vote weights be bigger, and a low VP lower to maximize VP usage. So really vote weight depends on pending curation efficiency VS efficiency i set as objective for the bot + current usage of VP. It's really only a matter of VP usage optimization. Vote weights would play a bigger role with much higher SP (like 500k+ as with that amount, you'll need to diversify your votes a lot more, so the weight efficiency arbitrage would have much more effect in that case)
For the selection, bot aims only at maximizing curation rewards, not quality or anything. I won't tell which criteria I use to determine which posts to vote on, nor at what time exactly, but the bot isn't really based on ML or anything. Just instructions so that he adapts all the time to select posts and optimize curation rewards at all times without any intervention needed. Let's say I learn and then change how the bot behaves via coding
That sounds really interesting, so you are using a rule based approach to detect posts and when to vote.
I haven't put much work into that aspect of my bot. It votes pretty much +- around the 5 minute mark.
But the whole thing about maximizing curation rewards sounds like an interesting reinforcement learning problem. Should be straight forward to learn an optimal behavior in regards to VP, time since post, number of people who have already voted etc. If I figure something out I'll do a post about that and let you know if you are interested :)
But building that by hand is much more impressive for sure :)
Posted using Partiko Android