Using　functional　connectivity　(FC)　or　effective　connectivity　(EC)　alone　cannot　effectively　delineate　brain　networks　based　on　functional　magnetic　resonance　imaging　(fMRI)　data,　limiting　the　understanding　of　the　mechanism　of　tinnitus　and　its　treatment.　Investigating　brain　FC　is　a　foundational　step　in　exploring　EC.　This　study　proposed　a　functionally　guided　EC　(FGEC)　method　based　on　reinforcement　learning　(FGECRL)　to　enhance　the　precision　of　identifying　EC　between　distinct　brain　regions.　An　actor-critic　framework　with　an　encoder-decoder　model　was　adopted　as　the　actor　network.　The　encoder　utilizes　a　transformer　model;　the　decoder　employs　a　bidirectional　long　short-term　memory　network　with　attention.　An　FGEC　network　was　constructed　for　the　enrolled　participants　per　fMRI　scan,　including　65　patients　with　tinnitus　and　28　control　participants　healthy　at　the　enrollment　time.　After　6　months　of　sound　therapy　for　tinnitus　and　prospective　follow-up,　fMRI　data　were　acquired　again　and　retrospectively　categorized　into　an　effective　group　(EG)　and　an　ineffective　group　(IG)　according　to　the　treatment　effect.　Compared　with　FC　and　EC,　the　FGECRL　method　demonstrated　better　accuracy　in　discriminating　between　different　groups,　highlighting　the　advantage　of　FGECRL　in　identifying　brain　network　features.　For　the　FGEC　network　of　the　EG　and　IG　per　state　(before　and　after　treatment)　and　healthy　controls,　effective　therapy　is　characterized　by　a　similar　pattern　of　FGEC　network　between　patients　with　tinnitus　after　treatment　and　healthy　controls.　Deactivated　information　output　in　the　motor　network,　somatosensory　network,　and　medioventral　occipital　cortex　may　biologically　indicate　effective　treatment.　The　maintenance　of　decreased　EC　in　the　primary　auditory　cortex　may　represent　a　failure　of　sound　therapy,　further　supporting　the　Bayesian　inference　theory　for　tinnitus　perception.　The　FGEC　network　can　provide　direct　evidence　for　the　mechanism　of　sound　therapy　in　patients　with　tinnitus　with　distinct　outcomes.