Learning from a reward signal rather than from labelled answers.
Colour is the family; a dashed line is the second member of it.
Learns the value of each state-action pair directly, converging to the optimal policy without ever modelling the transitions. · 1989
Replaces the Q table with a convolutional network reading pixels. The result that made reinforcement learning on raw sensory input look tractable. · 2013
Stores transitions in a buffer and samples them at random, so an update is not dominated by whatever just happened. · 2013
Blends TD residuals over many horizons with an exponential weight, giving one knob that trades bias against variance in the advantage estimate. · 2015
Constrains each update so the KL divergence between the old and new policy stays inside a trust region, giving a monotonic improvement guarantee at t… · 2015
Clips the ratio between new and old policy so an update cannot move too far from the data it was collected under. Most of TRPO's stability with none… · 2017
Samples a group of completions per prompt and uses their mean reward as the baseline, so the advantage is relative within the group. Drops the value… · 2024
Learns a policy and a value function together, using the value estimate as the baseline the policy gradient is measured against. Most modern policy m…
States, actions, transition probabilities and rewards, where the next state depends only on the present one. The formalism every method below is defi…
Differentiates expected return with respect to the policy parameters directly, so stochastic and continuous action spaces need no argmax over actions.
Updates a value estimate toward a later estimate rather than waiting for the final return. Learns from incomplete episodes, which is what makes onlin…