Roadmap¶

ArqonHPO follows semantic versioning. This document outlines our vision, planned features, and development priorities.

Current: v0.3.x (Stable)¶

Status: ✅ Released (January 2026)

The production-ready foundation for real-time hyperparameter optimization.

Core Capabilities¶

• PCR Algorithm — The Probe-Classify-Refine pipeline automatically selects strategies based on landscape structure
• Python Bindings — ArqonSolver for batch optimization, ArqonProbe for distributed LDS sampling
• CLI & Observability — TUI, web Dashboard, Prometheus metrics, structured logging
• Safety Executor — Constitutional guarantees with Guardrails, Rollback Policy, and audit trail
• 91% Test Coverage — Comprehensive unit and integration tests

What Makes v0.3 Production-Ready¶

Next: v0.4.0 (Q1 2026)¶

Status: 🔨 In Development

The "Infrastructure Phase" — making ArqonHPO deployable anywhere with comprehensive observability.

New Features¶

Adaptive Nelder-Mead with Restart Detection¶

The current Nelder-Mead implementation can stall when the simplex degenerates. v0.4 adds automatic restart detection:

// Detects when simplex volume collapses below threshold
// Automatically reinitializes with exploration phase
let nm = NelderMead::with_restart(config, restart_threshold);

• Problem: Simplex can collapse to a line in high dimensions
• Solution: Monitor simplex volume, trigger restart when degenerate
• Benefit: More robust convergence on ill-conditioned surfaces

Multi-Objective Optimization (MOO)¶

Support for Pareto-front optimization with multiple objectives:

config = {
    "objectives": ["latency", "throughput"],  # Both minimize
    "mode": "pareto"
}
solver = ArqonSolver(json.dumps(config))
candidates = solver.ask()  # Returns Pareto-optimal candidates

• Algorithm: NSGA-II style dominance sorting
• Output: Pareto front with configurable size
• Trade-off: User selects from Pareto set

ArqonShip v1.0 (Self-Healing CI)¶

The ArqonShip integration reaches v1.0 with:

• Automatic rollback on CI failure
• Bisection to find breaking commits
• Artifact correlation with test results
• Slack/Discord notifications

Infrastructure¶

Docker Images¶

Official multi-platform images:

docker pull ghcr.io/novelbytelabs/arqonhpo:0.4
docker pull ghcr.io/novelbytelabs/arqonhpo:0.4-slim  # ~20MB

Platforms: linux/amd64, linux/arm64, darwin/arm64

Helm Chart¶

Production Kubernetes deployment:

helm repo add arqon https://novelbytelabs.github.io/charts
helm install arqon arqon/arqonhpo \
  --set config.budget=1000 \
  --set monitoring.enabled=true

Includes: ServiceMonitor, PodDisruptionBudget, HPA

OpenTelemetry Integration¶

Full OTel support for distributed tracing:

arqonhpo --otel-endpoint http://otel-collector:4317 run ...

• Traces: ask/tell spans with parameter context
• Metrics: OTel metrics alongside Prometheus
• Logs: Structured logs with trace correlation

Future: v1.0.0 (Q2 2026)¶

Status: 🎯 Planned

The "Scale Phase" — GPU acceleration and distributed optimization for massive parameter spaces.

GPU-Accelerated Strategies¶

CUDA Backend¶

TPE kernel density estimation on GPU:

let tpe = TPE::cuda(device_id, dim);
// 100x speedup for high-dimensional KDE

• Speedup: 100x for dim > 100
• Memory: Configurable batch size
• Fallback: Graceful CPU fallback

Metal Backend (Apple Silicon)¶

Native M1/M2/M3 support via Metal:

solver = ArqonSolver(config, device="metal")

Distributed Optimization¶

Sharded Workers¶

Horizontal scaling across nodes:

# arqon-distributed.yaml
mode: distributed
coordinator:
  address: arqon-coordinator:9090
workers:
  count: 16
  strategy: round_robin
sharding:
  probe: by_index # ArqonProbe sharding
  refine: by_region # Spatial partitioning

• Coordinator: Central state, Raft consensus
• Workers: Stateless evaluators
• Communication: gRPC with protobuf

Fault Tolerance¶

Automatic recovery from worker failures:

• Checkpoint every N evaluations
• Replay from last checkpoint
• Worker health checks with timeout

Full Bayesian TPE¶

The current TPE uses point estimates. v1.0 adds full Bayesian treatment:

let tpe = TPE::bayesian(prior, acquisition_fn);

• Prior: Configurable Gaussian or Student-t
• Acquisition: EI, PI, UCB, Thompson Sampling
• Uncertainty: Credible intervals on predictions

Ecosystem Integrations¶

Long-Term Vision (v2.0+)¶

ArqonHPO as Infrastructure¶

Our north star: Optimization should be as reliable as a database.

Zero-Configuration Optimization¶

Fully automatic strategy selection:

# v2.0 vision: just specify the objective
solver = ArqonSolver.auto(budget=100)
# ArqonHPO infers bounds, selects strategy, tunes hyperparams

Sub-Microsecond Overhead¶

Target: <1μs latency for ask() in hot path:

Constitutional AI Safety¶

Extending constitutional safety to AI-in-the-loop:

constitution:
  max_regression_before_human: 3
  human_approval_required_for:
    - production_deploy
    - safety_override
  audit_retention: 90d

What We're NOT Building (Anti-Roadmap)¶

Clarity on scope:

Feature Requests & Voting¶

We prioritize based on community feedback:

1. GitHub Discussions — Propose new features
2. Vote on existing — 👍 reaction on issues
3. Contribute — PRs welcome for roadmap items

Version Comparison¶

Release Timeline¶

gantt
    title ArqonHPO Release Timeline
    dateFormat  YYYY-MM
    section Releases
    v0.3 (Current)      :done,    v03, 2026-01, 2026-01
    v0.4 Development    :active,  v04, 2026-01, 2026-03
    v0.4 Release        :milestone, 2026-03, 0d
    v1.0 Development    :         v10, 2026-03, 2026-06
    v1.0 Release        :milestone, 2026-06, 0d

Next Steps¶

• Changelog — Detailed version history
• Contributing — How to help
• About — Our philosophy