Batch vs Online Mode¶

ArqonHPO supports two primary modes of operation: Batch mode for offline experimentation with parallel evaluation, and Online mode for real-time, streaming optimization.

Mode Comparison¶

Batch Mode (PCR Workflow)¶

Batch mode uses the full Probe → Classify → Refine workflow. The solver requests multiple candidates at once, evaluates them (potentially in parallel), and reports all results together.

How It Works¶

1. ask() returns a batch of candidate configurations
2. You evaluate all candidates (can parallelize this step)
3. tell() reports all results back to the solver
4. Repeat until budget exhausted (returns None)

Example¶

from arqonhpo import ArqonSolver
import json

config = {
    "seed": 42,
    "budget": 100,
    "bounds": {"x": {"min": -5, "max": 5}, "y": {"min": -5, "max": 5}}
}

solver = ArqonSolver(json.dumps(config))
best = {"value": float('inf'), "params": None}

while True:
    # Get a BATCH of candidates
    batch = solver.ask()
    if batch is None:
        break

    # Evaluate all candidates (can parallelize!)
    results = []
    for i, params in enumerate(batch):
        value = objective(params)
        if value < best["value"]:
            best = {"value": value, "params": params}
        results.append({
            "eval_id": i,
            "params": params,
            "value": value,
            "cost": 1.0
        })

    solver.tell(json.dumps(results))

print(f"Best: {best}")

Phase Breakdown¶

The batch mode budget is divided across three phases:

The probe phase returns larger batches (to cover the search space), while the refine phase may return smaller batches depending on the selected strategy.

When To Use Batch Mode¶

• ✅ You can evaluate multiple candidates in parallel
• ✅ Full budget is known upfront
• ✅ Offline experiments (training runs, simulations)
• ✅ You want automatic strategy selection via PCR

Online Mode (Real-time Workflow)¶

Online mode bypasses the PCR workflow entirely. Each call returns exactly one candidate, and feedback is incorporated immediately before the next candidate is generated.

How It Works¶

1. ask_one() returns a single candidate configuration
2. You evaluate immediately
3. seed() reports the result back
4. Repeat until budget exhausted (returns None)

Example¶

from arqonhpo import ArqonSolver
import json

config = {
    "seed": 42,
    "budget": 100,
    "bounds": {"gain": {"min": 0.1, "max": 10.0}}
}

solver = ArqonSolver(json.dumps(config))

while True:
    # Get ONE candidate
    candidate = solver.ask_one()
    if candidate is None:
        break

    # Evaluate immediately
    value = evaluate_in_production(candidate)

    # Feed back immediately
    solver.seed(json.dumps([{
        "params": candidate,
        "value": value,
        "cost": 1.0
    }]))

Key Differences from Batch Mode¶

When To Use Online Mode¶

• ✅ Real-time control loops (adjusting parameters while system runs)
• ✅ Single evaluations at a time (can't parallelize)
• ✅ External evaluation systems (A/B testing, user feedback)
• ✅ Streaming data sources
• ✅ Continuous operation without defined budget

Choosing Between Modes¶

Decision Tree¶

graph TD
    A[Start] --> B{Can you parallelize<br>evaluations?}
    B -->|Yes| C{Known budget<br>upfront?}
    B -->|No| D{Real-time<br>constraint?}

    C -->|Yes| E[Use Batch Mode]
    C -->|No| F[Consider Batch Mode]

    D -->|Yes| G[Use Online Mode]
    D -->|No| H{Streaming<br>data?}

    H -->|Yes| G
    H -->|No| F

    style E fill:#22c55e,color:#fff
    style G fill:#6366f1,color:#fff

Practical Guidelines¶

Hybrid Approach¶

You can combine both modes. This is useful when you want the benefits of batch probing but need to switch to online refinement:

solver = ArqonSolver(json.dumps(config))

# Phase 1: Batch probing for landscape exploration
for _ in range(3):
    batch = solver.ask()
    if batch:
        results = parallel_evaluate(batch)
        solver.tell(json.dumps(results))

# Phase 2: Switch to online refinement
while True:
    candidate = solver.ask_one()
    if candidate is None:
        break
    value = evaluate_single(candidate)
    solver.seed(json.dumps([{
        "params": candidate,
        "value": value,
        "cost": 1.0
    }]))

This approach:

• Uses batch mode to efficiently explore the search space
• Switches to online mode for fine-grained, real-time refinement
• Maintains all history across both phases

API Reference¶

Batch Mode Methods¶

Online Mode Methods¶

Next Steps¶