Online Intelligent Parameter TuningPlatform for Industrial Fields
Turn experience-based parameter changes into a repeatable online experiment.
OptimLight is for intelligent online parameter tuning at industrial fields. It ships as general-purpose software and is not bound to one industry or one device.
Who it is for
- Site engineers install, start and stop, and uninstall.
- Tuning operators sign in, configure and run tunings, and read timeseries and history.
- Plant adaptation staff connect control write-out and time-series ingest from the point tables, then accept the adaptation.
What the platform does
Online adjustment
Connect to industrial fields and run online optimization in the real production environment.
Closed-loop iteration
Plan parameters → dispatch for execution → wait until stable → observe state, and keep iterating.
Automatic search
Use heuristic optimization algorithms to search the parameter space from the configured objectives and constraints.
Intelligent algorithms
Integrate Bayesian optimization, reinforcement learning (in development), and other advanced algorithms, significantly reducing field trials, for industrial cases where each evaluation is expensive.
General-purpose platform
Objectives, constraints, and controlled devices are configurable, without a process mechanism model.
Lightweight deployment
A B/S web architecture: simple to install and easy to deploy.
Online loop
A tuning iterates through plan, act, wait, and observe. After observe, if the task is not complete and the algorithm has not finished, it returns to plan.
Plan
From the last observation, the algorithm computes the next action. On the workbench you choose an optimizer type and configure the algorithm instances on that tuning.
Act
A set of device parameters is sent to the control interface. A live site uses the write-out path configured for that site. Bring-up and tests can use the built-in virtual plant.
Wait
After the write, the product first decides whether the machine is stable and the parameters have taken effect, then opens the measurement window. Without this step, observe would read a state that has not settled.
Observe
In the measurement window it reads machine state, computes current objective and constraint values, and evaluates them. Task completion is decided by the constraint evaluation in this step.
Platform layer and site adaptation layer
Each site runs the same platform-layer software: orchestration, tuning, measurement, query, supervision, and the UI. The platform does not embed which devices a site has.
Devices and points stay in the site adaptation layer. Control write-out and time-series ingest are configured per site, without changing platform code.
VirtualPlant is the built-in virtual plant. It replaces the physical process with a synthetic map for bring-up and test evaluation. The public demo will use this path and will not connect a real plant.
Optimizer types
In the UI you first choose an optimizer type, then configure algorithm instances on a tuning. Two types are current.
EA
Genetic / multi-objective evolutionary search. Use it when several objectives must be watched together and a population of candidate actions is searched.
BO
Bayesian optimization. Use it when each evaluation is expensive and the search must spend fewer steps. Reinforcement learning is in development.
Interface
- A tuning is a workspace. Create, copy, or open it from the list; configure and run it on the workbench.
- The workbench has four page tabs: Common, Algorithm, Monitor, and History. They belong to the same tuning, not four unrelated lists.
- The Timeseries page selects points, draws series, runs short-window trials, and exports. Charts for people use the query path and do not replace measurement in the loop.
- Manual control reads and writes by system or device, for comparison and for intervention when needed.
- Platform services list units on this machine, show observed state, and allow start and stop by permission. Supervision is not part of this loop.
Which sites it fits
The use-case pages name kinds of sites OptimLight can connect to, not customers. They cover continuous process setpoints, several objectives and constraints, expensive evaluations, and VirtualPlant first then site adaptation.