JAX PathwaysAll career routesCareer learning guide
RL engineer
Build experiments in which actions change later observations. You keep environment state, random sampling and policy updates reproducible, and evaluate improvement across seeds and conditions.
The work
Training reward rises while evaluation reward falls. Check leakage, environment termination, sampling and the evaluation policy.
Preparation
Be comfortable with a scalar objective, gradients, random keys and a training update. The route teaches functional environments and sampled objectives using a small finite environment.
Your portfolio
A policy-learning report with fixed evaluation conditions, seed variation and trajectory-level failure analysis.
Skills and evidence
Vectorized environments and PRNG
Required route practice
Check transitions, termination and owned random streams.
Explain the independent check, repeat it under a changed condition and retain the observed result.
Practice project: stages 1.
scan-based rollouts
Required route practice
Preserve active masks and return bookkeeping across time and environments.
Explain the independent check, repeat it under a changed condition and retain the observed result.
Practice project: stages 2.
Policy gradients and PPO
Required route practice
Check signed clipping cases and compare held-out returns with the exact evaluator.
Explain the independent check, repeat it under a changed condition and retain the observed result.
Practice project: stages 3, 4.
MLflow experiment provenance
Optional project extension; 1 project preparation lessons lie outside the required route
Keep training and evaluation seeds separate and retain complete run inventories.
Explain the independent check, repeat it under a changed condition and retain the observed result.
Practice project: stages 1, 2, 3, 4.
Milestones
Build the shared foundation
A reproducible regression fit with checked gradients and explicit random state.
Build a policy model
A reproducible classifier
Evaluate a learning agent
An evaluated policy in a functional environment
Review your readiness
Show that the evaluation measures policy improvement rather than a lucky seed or a changed environment.
- Are terminal transitions and padded steps handled correctly in return targets?
- How do results vary across trained seeds under the same held-out evaluation protocol?
- Does the policy-gradient estimate agree with the enumerated small-state reference?
- How do you compare logged policy runs without mixing training return and held-out evaluation?
This is a learning route for practicing the work, not a claim of job readiness. Use the milestone coverage and a reviewer’s feedback to identify what remains to demonstrate. The capability map below separates required route practice from optional extensions. Public checks and self-reported completion are not independent review or hardware qualification.