TD(λ)
A multi-step return method in temporal difference learning that uses a parameter λ to interpolate with exponential weighting between the one-step TD target and the Monte Carlo return, balancing bias and variance in long-horizon credit assignment.
Timeline (last 90 days)
FOCUS uses TD(λ) advantages to score segments between consecutive states, performing segment-level credit assignment combined with refined value estimates.
TD(λ) performs exponentially weighted interpolation between the one-step TD target and the Monte Carlo return via a parameter λ∈[0,1]; λ=0 degenerates to one-step TD, and λ=1 degenerates to the full-trajectory return.
Eligibility traces are the bridge connecting Monte Carlo and TD learning: a "trace" that decays over time records how recently each state was visited, and the λ parameter controls the degree of interpolation, making it the mechanistic basis of algorithms such as TD(λ) and SARSA(λ).
All Facts (3)
FOCUS uses TD(λ) advantages to score segments between consecutive states, performing segment-level credit assignment combined with refined value estimates.
50%UnverifiedTD(λ) performs exponentially weighted interpolation between the one-step TD target and the Monte Carlo return via a parameter λ∈[0,1]; λ=0 degenerates to one-step TD, and λ=1 degenerates to the full-trajectory return.
50%UnverifiedEligibility traces are the bridge connecting Monte Carlo and TD learning: a "trace" that decays over time records how recently each state was visited, and the λ parameter controls the degree of interpolation, making it the mechanistic basis of algorithms such as TD(λ) and SARSA(λ).
50%