The　past　few　years　have　witnessed　a　wide　deployment　of　low　earth　orbit　(LEO)　satellites　communications　and　networking.　With　the　explosive　growth　of　new　businesses,　satellite　network　is　expected　to　provide　global　coverage　and　high　bandwidth　availability　service.　Toward　this　end,　Multipath　TCP(MPTCP)　is　a　promising　transport　protocol　to　use　in　LEO　satellites　networks.　MPTCP　can　not　only　achieve　seamless　handover,　but　also　enhance　throughput　by　using　multiple　paths　transmission　mechanism.　However,　following　the　improvement　of　the　performance　and　scalability,　it　also　brings　unprecedented　challenges　for　congestion　control　of　multiple　sub-flows.　Especially,　currently　works　on　the　congestion　control　largely　relies　on　a　manual　process　which　presents　a　poor　performance　in　the　high-dynamic　complexity　network　environment.　Inspired　by　the　recent　success　of　applying　machine　learning　in　many　challenging　control　decision　domains,　such　as　video　game,　self-driving,　we　employ　deep　deterministic　policy　gradient　for　learning　the　optimal　congestion　control　strategies　by　interacting　with　the　underlying　network　environment.　Some　simulation　results　demonstrated　the　effectiveness　and　feasibility　of　our　architecture　and　algorithms.　©　2019　IEEE.