Multi-agent　decision　making　is　a　fundamental　problem　in　edge　intelligence.　In　this　paper,　we　study　this　problem　for　IoT　networks　under　the　distributed　Multi-Armed　Bandits　(MAB)　model.　Most　of　existing　works　for　distributed　MAB　demand　long-time　stable　networks　connected　by　powerful　devices　and　hence　may　not　be　suitable　for　mobile　IoT　networks　with　harsh　IoT　constraints.　To　meet　the　challenge　of　resource　constraints　in　mobile　IoT　environment,　we　propose　a　lightweight　and　robust　learning　algorithm　in　a　dynamic　network　allowing　topology　changes.　In　our　model,　each　agent　is　assumed　to　have　only　limited　memory　and　communicate　with　each　other　asynchronously.　Moreover,　we　assume　that　the　bandwidth　for　exchanging　information　is　limited　and　each　agent　can　transmit　O(log2K)　bits　(K　denotes　the　number　of　arms)　per　communication.　Rigorous　analysis　shows　that　despite　these　harsh　constraints,　the　best　arm/option　can　be　identified　collaboratively　by　the　agents　and　the　algorithm　converges　efficiently.　Extensive　experiments　illustrate　that　the　proposed　algorithm　exhibits　good　efficiency　and　stability　in　mobile　settings.　©　2020