Action Selection Methods
Do not treat convergence as an unconditional property of Semi-gradient Sarsa. The reported results depend on the action selection setting being discussed.
Why Action Choice Matters
The history of Semi-gradient Sarsa with function approximation is also a history of asking how action selection affects learning behavior. The method was first explored by Rummery and Niranjan in 1994. Later research examined whether particular combinations of function approximation and action selection would converge, and what convergence should mean in each setting.
Do not treat convergence as an unconditional property of Semi-gradient Sarsa. The reported results depend on the action selection setting being discussed.
From First Exploration to Later Studies
The historical sequence begins in 1994, when Rummery and Niranjan first explored Semi-gradient Sarsa with function approximation. Later, Gordon reported a specific convergence concern for linear Semi-gradient Sarsa using ε-greedy action selection. Still later, Precup and Perkins showed convergence in a differentiable action selection setting. These stages should be read as a progression of research questions rather than as one result that applies to every version of the method.
The Two Convergence Results
The ε-greedy result is deliberately qualified. For linear Semi-gradient Sarsa with ε-greedy action selection, Gordon reported that the method did not converge in the usual sense. However, it did enter a bounded region near the best solution. This is different from saying that learning becomes unbounded or that the method has no useful behavior.
The later differentiable-action-selection result is also conditional. Precup and Perkins showed convergence in a differentiable action selection setting. The source does not turn this result into a claim that every action selection rule produces convergence, nor does it provide enough detail here to generalize the result beyond the stated setting.
| Setting | Reported result | How to interpret it |
|---|---|---|
| Linear Semi-gradient Sarsa with ε-greedy action selection | Does not converge in the usual sense; enters a bounded region near the best solution | A bounded-region result is not the same as ordinary convergence |
| Differentiable action selection | Convergence was shown | The result is tied to the differentiable action selection setting |
The two results are conditional on their stated action selection settings.
Connecting the Four Ideas
Four ideas must be kept together when reading this research history: Semi-gradient Sarsa is the learning method, function approximation describes how the method represents what it learns, action selection specifies the setting in which actions are chosen, and convergence is the behavior being studied. Changing the action selection setting changes which convergence statement can be supported.
Reading a conditional claim
Suppose a research note says that one action selection setting gives a bounded-region result and another setting gives convergence. What should a careful reader conclude?
- 1Identify the method
Recognize that both statements concern Semi-gradient Sarsa with function approximation.
- 2Identify the action selection setting
Keep the ε-greedy and differentiable settings separate instead of merging them into one general case.
- 3Match each outcome
Associate the bounded-region result with linear Semi-gradient Sarsa using ε-greedy action selection, and associate the convergence result with the differentiable action selection setting.
- 4Avoid overgeneralizing
Do not conclude that Semi-gradient Sarsa always converges or never converges. The evidence supports conditional claims.
The correct interpretation is setting-specific: the reported outcome depends on the action selection rule and the other conditions named in the result.
Mistakes in Reading the History
Common mistake Saying that linear Semi-gradient Sarsa with ε-greedy action selection converges.
What it looks likeThe method converges normally.
Why it failsThe reported result says that it does not converge in the usual sense, although it enters a bounded region near the best solution.
Do this insteadDescribe the bounded-region result and distinguish it from ordinary convergence.
Common mistake Saying that Semi-gradient Sarsa never converges.
What it looks likeNo version of the method can converge.
Why it failsConvergence was later shown in a differentiable action selection setting.
Do this insteadState the action selection setting before making a convergence claim.
Common mistake Treating the differentiable-action-selection result as applying to ε-greedy action selection.
What it looks likeBecause convergence was shown for differentiable action selection, it also follows for ε-greedy selection.
Why it failsThe source reports the results as belonging to different settings.
Do this insteadKeep the two settings and their reported outcomes separate.
Common mistake Using an application reference to supply missing convergence details.
What it looks likeInferring a convergence result from an application mentioned elsewhere.
Why it failsThe source states that convergence claims here come from the explicitly reported action selection settings, not from an unstated interpretation of an application reference.
Do this insteadBase the claim on the named action selection setting and reported result.
When summarizing a convergence result, name the method, the function-approximation condition, the action selection setting, and the exact strength of the reported conclusion. This prevents a conditional result from becoming an unconditional claim.
Check Your Interpretation
Write a two-sentence comparison of the ε-greedy and differentiable action selection results. Your first sentence should state the reported outcome for linear Semi-gradient Sarsa with ε-greedy action selection. Your second sentence should state the later result for a differentiable action selection setting without extending it to other settings.
Hints
- Use the phrase bounded region near the best solution for the ε-greedy result.
- Use the phrase convergence was shown for the differentiable setting.
- Mention that the claims are conditional on their action selection settings.
What do you think happens?
A result says that a method enters a bounded region near the best solution but does not converge in the usual sense. Should you summarize that as convergence?
Reveal answer
No, because the source distinguishes the bounded-region result from convergence in the usual sense.
Precise research reading requires preserving the distinction between the reported bounded-region behavior and the later convergence result in a differentiable action selection setting.
Historical Takeaway
The research history moves from initial exploration to increasingly specific convergence questions. Rummery and Niranjan first explored Semi-gradient Sarsa with function approximation in 1994. Gordon later reported that linear Semi-gradient Sarsa with ε-greedy action selection does not converge in the usual sense, while entering a bounded region near the best solution. Precup and Perkins later showed convergence in a differentiable action selection setting. The main lesson is to attach every convergence claim to its stated action selection conditions.
At a glance
- Semi-gradient Sarsa with function approximation was first explored by Rummery and Niranjan in 1994.
- Linear Semi-gradient Sarsa with ε-greedy action selection was reported not to converge in the usual sense, although it entered a bounded region near the best solution.
- Convergence was later shown in a differentiable action selection setting.
- The evidence supports conditional, setting-specific convergence claims rather than an unconditional claim that the method always converges or never converges.