Optimization Track¶
This track focuses on improving model and policy quality under resource and feedback constraints.
Recommended Sequence¶
projects/automl-hpo-showcaseprojects/autoresearchprojects/agentic-course-assistant-showcaseprojects/rl-bandits-policy-showcaseprojects/student-support-rl-showcaseprojects/sota-supervised-learning-showcase(benchmark extension)
Core Skills Covered¶
- Hyperparameter search strategy tradeoffs.
- Fixed-budget experiment governance for agent-authored changes.
- Prompt and platform tradeoffs for autonomous research workflows.
- Agent routing, tool use, guardrails, and trace interpretation.
- Experiment logging and comparability.
- Bandit policy reward/regret analysis.
- Contextual bandits, MDP design, Bellman intuition, tabular Q-learning, and the DQN/PPO bridge.
- Reward hacking diagnosis and offline policy-governance tradeoffs.
- Choosing optimization objectives aligned with business constraints.
Evidence Artifacts To Inspect¶
artifacts/hpo/trials.csvartifacts/hpo/strategy_comparison.csvprojects/autoresearch/artifacts/overview/platform_comparison.csvprojects/autoresearch/artifacts/analysis/decision_scenarios.csvprojects/agentic-course-assistant-showcase/artifacts/agent_trace.jsonprojects/student-support-rl-showcase/artifacts/q_learning/training_curve.csvprojects/student-support-rl-showcase/artifacts/eval/policy_comparison.csv- reward/regret outputs in RL bandits artifacts
- experiment logs in
artifacts/experiments/
Suggested Reflection Prompts¶
- Which search strategy wins under strict runtime budget?
- When is a tiny improvement too expensive in code complexity?
- How should the same research loop change between Apple Silicon and CUDA?
- Which agent decisions should be deterministic before a hosted model is introduced?
- How sensitive are recommendations to objective choice?
- What offline checks are required before online policy rollout?
- Which reward terms create the biggest mismatch between local reward and true objective?