Distributed　Traffic　Engineering　(TE)　adopts　decentralized　routing　and　offers　natural　merits　over　centralized　TE　in　rapidly　responding　to　dynamic　network　flows.　However,　based　solely　on　local　network　observations,　traditional　distributed　TE　methods　often　lack　sufficient　routing　evidence　necessary　for　generating　optimal　policies.　Furthermore,　reinforcement　learning,　commonly　used　in　distributed　TE　methods　to　avoid　the　need　for　labeled　data,　faces　challenges　in　accumulating　valuable　routing　experiences　due　to　its　trial-and-error　exploration.　To　handle　these　challenges,　we　propose　a　novel　Distributed　Intelligent　Traffic　Engineering　(DITE)　framework　that　empowers　geographically　distributed　agents　to　rapidly　and　adaptively　optimize　routing　policies　for　dynamic　network　flows　with　low　communication　overhead.　Specifically,　to　address　the　limitation　of　optimizing　local　routing　based　solely　on　partial　network　observations,　we　propose　enriching　the　routing　evidence　for　each　agent　with　sparse　network-wide　information.　To　promote　agents　to　learn　cooperative　routing　polices　efficiently,　we　explore　a　new　paradigm　for　training　routing　agents　by　seamlessly　integrating　imitation　learning　with　reinforcement　learning.　Additionally,　to　efficiently　describe　the　implicit　spatiotemporal　features　among　the　routing　evidence,　we　customize　a　transformer-based　agent　to　utilize　the　attention　mechanism　for　establishing　the　relationships　between　agent　states　and　routing　policies.　Extensive　comparative　evaluations　on　three　real　network　topologies　and　traffic　traces　demonstrate　the　superior　performance　of　DITE　in　achieving　distributed　TE.　Compared　to　the　existing　distributed　TE　approaches,　DITE　achieves　a　notable　reduction　of　35.95%–42.35%　in　maximum　link　utilization　and　a　reduction　of　23.62%-68.39%　in　end-to-end　latency.　Moreover,　the　experimental　results　indicate　DITE　is　robust　to　network　failures　and　traffic　changes.　©　2013　IEEE.