Gate chunking on RL convergence (EMA of |delta_Q|)#577
Draft
kimjune01 wants to merge 4 commits intoSoarGroup:developmentfrom
Draft
Gate chunking on RL convergence (EMA of |delta_Q|)#577kimjune01 wants to merge 4 commits intoSoarGroup:developmentfrom
kimjune01 wants to merge 4 commits intoSoarGroup:developmentfrom
Conversation
Gate chunking on RL convergence by tracking an exponential moving average (EMA) of |delta_Q| per production rule. When all RL rules contributing numeric-indifferent preferences to a slot have converged (EMA below threshold), the decision is made greedily instead of stochastically. This makes the decision deterministic, which enables chunking to compile the converged policy into a production rule. New parameters (all under rl): chunk-gate on/off (default off) — enable convergence gating chunk-gate-threshold double (default 0.01) — EMA below this = converged chunk-gate-ema-decay double (default 0.95) — EMA smoothing factor When chunk-gate is off, behavior is identical to the existing codebase. Motivation: Laird (2022) §4 identifies the RL–chunking composition gap as a known limitation. RL uses stochastic exploration, chunking requires deterministic results, so the two cannot compose. The planned fix is to gate chunking on RL convergence. This patch implements that gate. Reference: "Introduction to the Soar Cognitive Architecture" (Laird, 2022, arXiv:2205.03854), §4, p.10.
Three FullTests covering the chunk-gate feature: 1. testRLConvergenceGate: agent with RL learning and chunk-gate ON (fast EMA decay=0.5, threshold=0.1). Two operators with consistent reward. Verifies 50 decisions complete without crash/hang. 2. testRLConvergenceGateOff: same agent with chunk-gate OFF. Regression test — identical decision count confirms no behavior change when feature is disabled. 3. testRLConvergenceGateParams: verifies the three new parameters (chunk-gate, chunk-gate-threshold, chunk-gate-ema-decay) are accepted by the command parser with valid values.
Three FullTests covering the chunk-gate feature: 1. testRLConvergenceGate: agent with RL learning and chunk-gate ON (fast EMA decay=0.5, threshold=0.1). Two operators with consistent reward. Verifies 50 decisions complete successfully. 2. testRLConvergenceGateOff: same agent with chunk-gate OFF. Regression test — identical decision count confirms no behavior change when feature is disabled. 3. testRLConvergenceGateParams: verifies the three new parameters (chunk-gate, chunk-gate-threshold, chunk-gate-ema-decay) are accepted by the command parser with valid values. All three tests pass. Existing RL/chunking tests (Chunk_RL_Proposal, RL_Variablization, testPreferenceSemantics, testLearn) also pass, confirming zero regression.
The two test agents were nearly identical — only the chunk-gate params differed. Now there's one shared agent file; the C++ tests set rl params via ExecuteCommandLine before sourcing.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Experimental / demonstration only. This PR implements a convergence gate for RL-chunking composition as described in Diagnosis: Soar and Prescription: Soar. It is intended to illustrate the approach and invite discussion, not for immediate merge.
chunk-gate(on/off),chunk-gate-threshold(default 0.01),chunk-gate-ema-decay(default 0.95)Motivation
Laird (2022) §4 identifies the RL–chunking composition gap: "RL uses stochastic selection, chunking requires deterministic results, so the two cannot compose." The planned fix is gating chunking on RL convergence. This patch implements that gate.
The same EMA that triggers compilation can later trigger invalidation (detecting divergence after a period of stability), but that is left for a follow-up.
Changes
production.hrl_ema_delta_qfield to production structproduction.cpp,rete.cpp,reinforcement_learning.cppreinforcement_learning.hrl_slot_converged()reinforcement_learning.cpprl_perform_update(), add param init, implementrl_slot_converged()decide.cpprun_preference_semantics(), check convergence before exploration policy; if converged, select greedilyTests
All pass locally (M4 Pro, macOS, 16s clean build):
testRLConvergenceGate— agent with chunk-gate ON, 50 decisions, completes successfullytestRLConvergenceGateOff— same agent with chunk-gate OFF, regression testtestRLConvergenceGateParams— verifies new params accepted by command parserChunk_RL_Proposal,RL_Variablization,testPreferenceSemantics,testLearn) all passReference
See also: #578 — Episodic-to-semantic consolidation and eviction (prescription steps 1 & 2)