Use this reinforcement-learning update to study policy evaluation or gradient-based learning in a fully specified environment with explicit states, actions, rewards, discounting, and sampling behavior.

Policy logits

Required list input.

Selected action

Required integer input.

Episode return

Required number input.

Baseline

Required number input.

Learning rate

Required number input.

Apply one categorical REINFORCE policy-gradient update from an action, return, and baseline. The tool evaluates the supplied inputs together and returns the named outputs below; it does not infer omitted operating conditions or change the units shown.^[1]

Policy update

The resulting policy update returned as an object.

• Estimates can be biased or high-variance and depend on exploration, bootstrapping, function approximation, rollout length, advantage normalization, off-policy corrections, seeds, and environment stationarity. Training return does not establish safe deployment.
• Use finite inputs in the displayed units and preserve more precision than the final presentation requires. Independently verify safety-critical, financial, compliance, or production decisions.

Test tabular cases, run multiple seeds, report confidence intervals and sample cost, evaluate off-policy and under perturbations, and apply domain safety constraints.

• A3C Rollout Update Calculator

  Calculate discounted returns, advantages, and A3C rollout losses.

• Reinforcement-Learning Policy Evaluator

  Evaluate a fixed finite-state policy by iterating rewards and its transition matrix.

• Gradient Descent Optimizer Step

  Apply one gradient-descent update x − η∇f; the caller supplies the gradient and learning rate.