Positive Reinforcement During Computer Analysis

I really appreciate the discussion!

Sybotes, Toadofsky, if you have links to previous posts here or elsewhere regarding this I'd love to take a look.

jperkins that article is definitely more towards what I'm thinking about. I started looking more into some scientific supporting arguments, but I'm not cognitive psychologist/learning expert. This Wiki article was fun to start reading: https://en.wikipedia.org/wiki/Operant_conditioning
My brief understanding of that and a few published article summaries seems to suggest that a mix of both positive and negative reinforcement is ideal for building behavior/learning.

I'm not sure that it would be more resource-intensive then what already is being computed. A "best (least bad) move" is already being computed during analysis, I think just displaying something like "(#total_moves - (#inaccuracies + #mistakes + #blunders))/(#total_moves)" could be simple positive feedback.

There's definitely an interesting discussion about where the line is between "what lichess should do" and "what you should get elsewhere". I would argue that if mistakes/centipawn loss/etc. are provided, then any positive metrics that can be computed in a similar manner fall into the same category.

I'll attempt to spend some time looking at the front-end code and see if there's anything simple I can come up with. Again the discussion is much appreciated!

What about some kind of "criticism sandwich" algorithm? Wedge the criticism between two loaves of "compliment" bread.

[compliment] [criticism] [compliment2]

I'm not a programmer, but more of an amateur software architect, and that was me diagramming my idea.

Perhaps some random selection of compliment from a compliment database? Then just insert the normal LiChess Stockfish commentary. Followed by another compliment, or even create an encouraging remark database?

But in all seriousness, I really appreciate the positive thinking from your post. I understand that it's a bit of computer science/cognitive science challenge to code this without sending out too many false positives (no pun intended). But it's great ideas like this that bring change. Thank you.

Jeff

#12 "But it's great ideas like this that bring change."

Unfortunately the world is not quite that simple... sorry. Great ideas which can't be executed can't be used; and in software details matter quite a bit.
http://xyproblem.info/

Do you really need Stockfish to stoke your ego? When I analyse with a computer what I want to know first and foremost is where I can improve.

№ 13,

Well, the entire Modern era is an example of nice-sounding ideas that have already been proven not to work at all, under any conditions whatsoever, still being tested ad nauseam anyway by those in power. 😑 (100 million deaths … 200 million deaths … can we get 300 million?? … SOLD to the highest voter!)

I know it’s a pessimistic thing to say, but … just saying.

PS: I wish those in power in “the real world” were all programmers; or at least intelligent, honest and reasonable like you. 🧡

I regret not being more direct about my implementation ideas in the first post, regarding simply showing inverse statistics, etc. In hindsight I understand it comes off a bit like $OBLIVIOUS_USER (#13).

I do not think that comments like #14 (and the jokes in #12) are very productive. Is your argument that "people shouldn't be babies" ? If this is the majority opinion I'd be disappointed; I very much love lichess and its vision but would not be nearly as excited to engage if suggestions for beginner accessibility and improvement were dismissed by "don't be a snowflake" at every turn. The engine is already telling you where you went wrong, I see no logical reason for why this is any different from telling you where you went right.

From my brief research, there seems to be a decent amount of scientific evidence that positive reinforcement is valuable for learning. Skimming these articles suggests a mix may be best for maximizing reward (in this case learning).
https://pubs.aeaweb.org/doi/pdf/10.1257/000282803322157142 (direct PDF)
https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1006973 (journal article landing page)

If someone has evidence to the contrary I would love to see it, and am open to my opinion on this being swayed. I'm less likely to be convinced by coding difficulty arguments; but if that's the case I'll work on the dev onboarding and see what I can come up with.