State, Behavior, Feedback & Adaptation

A system's state is the information needed at one instant so that, together with future inputs, the model can determine future behavior. It is not every fact about the past and not necessarily every stored variable. State is relative to the model and prediction task.

For a discrete-time model,

$$x_{k+1}=f(x_k,u_k,d_k), \qquad y_k=g(x_k,u_k),$$

where $x_k$ is state, $u_k$ a chosen input, $d_k$ a disturbance and $y_k$ an observation. A continuous-time model replaces the update with $\dot{x}=f(x,u,d)$. A hybrid system combines continuous evolution with discrete events such as switching, failure, admission or mode changes.

State is a compression of history #

The state must make the future conditionally independent of earlier history under the model. Position alone is not a sufficient state for a moving mass because velocity affects what happens next. A TCP connection state includes sequence and acknowledgement information because the last packet alone does not determine the next legal transition. A queue length may be sufficient for a memoryless service model and insufficient when jobs have different remaining sizes.

Choosing too little state produces unexplained residual dependence. Choosing too much creates an unidentifiable, expensive model. The test is predictive: after conditioning on proposed state and inputs, does earlier history still improve predictions in a repeatable way?

Behavior is a set of possible trajectories #

One observed trace is not “the behavior.” Initial conditions, inputs, disturbances and parameter uncertainty define a family of trajectories. Important qualitative structures include:

• equilibria, where state stops changing under fixed inputs;
• cycles, where trajectories repeat;
• transients, the path between regimes;
• attractors and basins, which describe long-run destinations and the initial states leading to them;
• invariants, properties preserved by every allowed transition;
• reachable and forbidden states, central to both design and safety.

Positive feedback can amplify deviation or create multiple stable regimes. Negative feedback can reject disturbances and stabilize a target, but delay, saturation and wrong sign can make the same loop oscillate or diverge. “Feedback” alone says nothing about stability.

Change, development and lifecycle #

Separate three timescales:

1. operation: state changes under a fixed structure and law;
2. adaptation: parameters, policies or learned estimates change in response to experience;
3. evolution or development: structure, purpose, environment or selection pressures change.

A thermostat cycling a heater is operating. A controller estimating plant gain online is adapting. Replacing the heating plant and changing the building's occupancy model is system evolution. Confusing these levels makes validation impossible: evidence for a fixed controller does not validate its learning rule, and evidence for the learning rule does not validate unbounded redesign.

flowchart TD
  E["Environment and disturbances"] --> S["System state"]
  S --> Y["Observed behavior"]
  Y --> C["Feedback policy"]
  C --> S
  Y -.-> L["Adaptation law"]
  L -.->|"changes policy or model"| C

An adaptive loop creates a second feedback loop. It needs its own stability, excitation, constraints and rollback reasoning.

Self-organization without mysticism #

Self-organization means macroscopic order develops through local interactions without a single agent specifying the final configuration. It does not mean “without rules,” “without energy,” or “without boundary conditions.” Synchronization, clustering and load balancing can be self-organized when local coupling produces global structure.

To analyze such a claim, specify the local update rule, conserved or supplied resources, noise, topology and order parameter used to measure the global pattern. Then test whether the pattern persists across initial conditions and whether intervention on the proposed coupling changes it.

A practical behavior model #

For any system under study, record:

• state variables and why they are sufficient;
• inputs, disturbances and observations;
• update/flow equations or transition rules;
• modes and mode-switch conditions;
• constraints and invariants;
• equilibria, transients and unsafe regions;
• the timescale on which parameters or structure may change.

This turns vague statements such as “the system learns,” “the process stabilizes,” or “the organization adapts” into claims that can be simulated, measured and falsified.