"in the free play section we need a 'for you' section that recommends pieces to the player based on what pieces they like to play, what pieces they have completed (showing their level) and popular pieces that might be a good fit for them. come up with 5 different algorithms in the lab and show me their results for each"
Every number comes from a frozen snapshot of the real game: – players, – runs, – public pieces, with bots, admins and our own network taken out the way the Popular shelf takes them out. Players are pseudonyms (P001 is the one who has played most). Pick a player on the left and every algorithm runs on their history live; the scoreboard runs them all on everyone.
The backtest: for each real player, hide a piece they went on to play, run each algorithm on the history before it, and ask whether that piece was on the shelf of 8. – cases, up to ten a player. It favours popularity a little: what people chose next was steered by the guides and the Popular shelf the game already had.
"we need next to each piece a 'progress bar'. so players will feel like they are advancing completion everytime they play a piece. but how do we measure progress? come up with 5 different progress calculation templates that's based simultaneously on results + work (perhaps we can have 2 separate progress bars, one for progress and one for results/skill)"
Practice (work) moves when you play. Skill (results) moves when you play it better. Neither ever goes down: a bad run adds practice and takes nothing away. Four templates draw the two as two bars; the fifth folds them into one bar whose ceiling is set by the results.
Each of – real (player, piece) histories replayed run by run. Moves counts runs where a bar the player sees moved at least a point, while there was still something to earn and the run's end is known.