Introducing CLP: coupled learning and planning from limited experience and delayed outcomes.
@rei_labs16K followers
13,367 views
25 Comments
Comments from the original post on X (Twitter)

@rei_labs
We’ve documented our exploration of how retained values and routines make later planning more useful, and how planning supports further learning. The full article includes the experiments and results.
38 likes

@rei_labs
Training data is finite. Progress may require experience that does not exist yet. Adapt-1, our general learner, uses retained values and routines to guide trials, then learns from planning assessments and observed outcomes without demonstrations or expert action labels.
31 likes

@rei_labs
The chess experiment began with 800 games. Records contained twelve piece counts and an eventual outcome; 77.1% followed random moves. No expert move labels or engine evaluations were supplied. The first fit had little direct guidance about good decisions.
18 likes

@rei_labs
After an initial value fit, one-ply lookahead reconsidered stored positions. Search assessments helped form later targets, while observed outcomes kept a separate contribution. Revised values then changed decisions under the same planning settings.
19 likes

@rei_labs
In Machina, search compares four complete eight-command routines. The selected routine runs separately; its commands and measured outcome enter learning. Each acquisition run generated 2,048 proposals and admitted 512 executions.
19 likes

@rei_labs
Four candidates per choice can matter across hundreds of rounds: a selected execution changes the basis of later proposals. In the recorded trace, a new parent changed all four next routines, even with the same requested seed.
18 likes

@rei_labs
The result is a retained skill: an acquired routine plus a learned correction for the starting observation. Both stay fixed during use, while commands vary with the start. The skill can execute without another search.
19 likes

@rei_labs
After acquisition and refinement, the frozen docking skill succeeded on 34/64 new starts through direct execution, or 52/64 with evaluation search. No learning took place during evaluation. API details and examples:
24 likes

@0xreitern
crude learning: learning from scratch without demonstration, engine label or evaluation retained learning: success in applying previously learned skill in new initial environment
14 likes

@guid_o
look at me, i'm reitard now!
9 likes

@Ruimtepak_
greitness
7 likes

@ifexito1
LFG More gigabrain tech I’ll never fully comprehend
6 likes

@0xacrnm
technically speaking the data you provided adapt-1 is not just noisy it’s bad but it still got something out of it
6 likes

@kn0xhound
👀👀👀
6 likes

@Lojom53
Interesting direction. Coupling learning and planning via retained state makes sense with scarce experience and delayed feedback. Chess setup is clean: piece counts + outcomes (mostly random), then lookahead improves targets. Curious how far this scales. Grei
6 likes

@8480Howard
Grei
5 likes

@abdominoux
Tech is Greit
5 likes

@moreir92
Grei
2 likes

@kaliXBT
👀
1 like

@Hovavayo
Hey @grok sıfırdan yaradıblar yoxsa fine tune ediblər
1 like




