This　paper　presents　a　novel　multi-agent　deep　reinforcement　learning　(MADRL)　approach　for　real-time　rescheduling　of　rail　transit　services　with　short-turnings　during　a　complete　track　blockage　on　a　double-track　service　corridor.　The　optimization　problem　is　modeled　as　a　Markov　decision　process　with　multiple　control　agents　rescheduling　train　services　on　each　directional　line　for　system　recovery.　To　ensure　computational　efficacy,　we　employ　a　multi-agent　policy　optimization　solution　framework　in　which　each　control　agent　employs　a　decentralized　policy　function　for　deriving　local　decisions　and　a　centralized　value　function　approximation　(VFA)　estimating　global　system　state　values.　Both　the　policy　functions　and　VFAs　are　represented　by　multi-layer　artificial　neural　networks　(ANNs).　A　multi-agent　proximal　policy　optimization　gradient　algorithm　is　developed　for　training　the　policies　and　VFAs　through　iterative　simulated　system　transitions.　The　proposed　framework　is　implemented　and　tested　with　real-world　scenarios　with　data　collected　from　London　Underground,　UK.　Computational　results　demonstrate　the　superiority　of　the　developed　framework　in　computational　effectiveness　compared　with　previous　distributed　control　algorithms　and　conventional　metaheuristic　methods.　We　also　provide　managerial　implications　for　train　rescheduling　during　disruptions　with　different　durations,　locations,　and　passenger　behaviors.　Additional　experiments　show　the　scalability　of　the　proposed　MADRL　framework　in　managing　disruptions　with　uncertain　durations　with　a　generalized　model.　This　study　contributes　to　real-time　rail　transit　management　with　innovative　control　and　optimization　techniques.