The　integration　of　distributed　generations　(DG),　such　as　wind　turbines　and　photovoltaics,　has　a　significant　impact　on　the　security,　stability,　and　economy　of　the　distribution　network　due　to　the　randomness　and　fluctuations　of　DG　output.　Dynamic　distribution　network　reconfiguration　(DNR)　technology　has　the　potential　to　mitigate　this　problem　effectively.　However,　due　to　the　non-convex　and　nonlinear　characteristics　of　the　DNR　model,　traditional　mathematical　optimization　algorithms　face　speed　challenges,　and　heuristic　algorithms　struggle　with　both　speed　and　accuracy.　These　problems　hinder　the　effective　control　of　existing　distribution　networks.　To　address　these　challenges,　an　active　distribution　network　dynamic　reconfiguration　approach　based　on　an　improved　multi-agent　deep　deterministic　policy　gradient　(MADDPG)　is　proposed.　Firstly,　taking　into　account　the　uncertainties　of　load　and　DG,　a　dynamic　DNR　stochastic　mathematical　model　is　constructed.　Next,　the　concept　of　fundamental　loops　(FLs)　is　defined　and　the　coding　method　based　on　loop-coding　is　adopted　for　MADDPG　action　space.　Then,　the　agents　with　actor　and　critic　networks　are　equipped　in　each　FL　to　real-time　control　network　topology.　Subsequently,　a　MADDPG　framework　for　dynamic　DNR　is　constructed.　Finally,　simulations　are　conducted　on　an　improved　IEEE　33-bus　power　system　to　validate　the　superiority　of　MADDPG.　The　results　demonstrate　that　MADDPG　has　a　shorter　calculation　time　than　the　heuristic　algorithm　and　mathematical　optimization　algorithm,　which　is　useful　for　real-time　control　of　DNR.　©　2019　Power　System　Protection　and　Control　Press.