In　the　present　study,　brain　activation　patterns　of　heuristic　problem　solving　were　investigated　in　the　context　of　the　puzzle　Sudoku　experiment　by　using　a　two-stage　clustering　approach.　The　cognitive　experiment　was　composed　of　easy　tasks　and　difficult　tasks.　In　the　two-stage　clustering　approach,　K-means　served　as　the　data　selection　role　in　the　first　stage　and　the　affinity　propagation　(AP)　served　as　partition　role　in　the　second　stage.　Functional　magnetic　resonance　imaging　(fMRI)　was　used　to　collect　the　slow　event　related　paradigm　data.　Simulated　fMRI　datasets　were　used　to　evaluate　the　validity　of　the　clustering　method　and　compare　the　performance　of　fuzzy　c-means　(FCM)　as　an　alternate　method　in　the　first　stage.　Test　results　illustrated　that　the　performance　of　K-means　in　this　role　was　better　than　that　of　FCM.　Further,　the　proposed　method　was　applied　to　the　heuristic　problem　solving　fMRI　data　and　the　results　showed　that　the　brain　activation　patterns　observed　in　the　experiment　exhibited　compact　and　coherent　activity　mode　in　dealing　with　different　cognitive　tasks.