Skip to content

01_001_what_state_collapser_is.md

Latest commit

 

History

History
78 lines (58 loc) · 2.53 KB

01_001_what_state_collapser_is.md

File metadata and controls

78 lines (58 loc) · 2.53 KB
Raw

What state_collapser Is

state_collapser is a structural layer for reinforcement-learning problems.

It is easiest to place it next to familiar RL tools by contrast:

RLlib says:
    Give me an env; I will run scalable RL algorithms on it.

Stable-Baselines3 says:
    Give me an env; I will run standard reliable RL algorithms on it.

state_collapser says:
    Give me an env or discovered transition system; I will construct a better
    hierarchical/quotient decision structure around it.

The package is therefore not trying to be a general training framework. It does not currently own neural model families, optimizers, distributed rollout, replay buffers, checkpoints, or experiment manifests. Those belong to the engineer, the learner, or a later adapter.

The package owns a different layer:

environment or transition system
    -> discovered graph
        -> partition tower
            -> quotient/fiber decision structure
                -> learner-facing inputs and transitions

That layer is not limited to RL. A downstream graph-ML package can also hand state_collapser a known graph, treat it as already discovered, build a partition tower, and lift coarse computations back along node and edge fibers. The first concrete downstream application of this form is HGraphML.

In short:

The package builds the stage.
You bring the learner.
The fiber is the lift.

Current Maturity

This is still pre-alpha research-mode infrastructure. It has real runtime surfaces and tests, but it should not be mistaken for a mature RL framework.

What exists now:

  • graph, vista, and tower runtime machinery
  • persistent state/action partition towers
  • tower-aware training inputs, decisions, transitions, collectors, and reference loops
  • FrozenQuotientBehavior, PathFiber, and FiberConditionedStage
  • backend-independent linearization records, shared tower encodings, benchmark mode reports, and optional Torch batch conversion
  • example environments and smoke training paths

What does not exist yet:

  • serious PyTorch model families
  • vectorized rollout
  • full replay/checkpoint/manifest infrastructure
  • RLlib or Stable-Baselines3 adapters
  • production-scale benchmarking claims

For the implemented training-stage surface, start with fiber-conditioned training.

For the current tensor/model boundary, read tensorization boundary.

For the first non-RL downstream application, read downstream applications.