Currently,　congestion　has　become　a　common　urban　disease.　Recently,　traffic　control　methods　leveraging　reinforcement　learning　(RL)　have　gained　significant　attention　due　to　their　capability　to　effectively　utilize　real-time　traffic　data.　However,　existing　RL-based　methods　for　multi-intersection　traffic　signal　control　(TSC)　solely　focus　on　enhancing　network-level　performance,　which　can　potentially　exacerbate　traffic　conditions　at　specific　intersections,　making　it　unreasonable.　This　balance　of　network-level　and　local　intersection　traffic　conditions　is　referred　to　as　the　fairness　issue　in　TSC.　To　address　the　aforementioned　fairness　issue　and　further　alleviate　congestion,　a　multi-agent　reinforcement　learning　(MARL)　algorithm,　COunterfactual　Multi-Agent　policy　gradients　(COMA),　which,　compared　to　Multi-agent　Advantage　Actor-Critic　(MA2C),　is　characterized　by　its　use　of　counterfactual　advantage　estimation　to　effectively　resolve　the　credit　assignment　problem,　was　employed　in　multi-intersection　TSC　domain.　However,　as　an　on-policy　algorithm,　COMA　encounters　limitations　in　training　sample　utilization,　specifically　in　terms　of　inefficient　sample　usage　and　temporal　correlation　among　samples.　To　enhance　the　efficiency　of　sample　usage　and　to　mitigate　the　temporal　correlation　among　training　samples,　thereby　improving　overall　training　efficiency,　we　introduced　(a)　importance　sampling　and　(b)　distributed　computing.　COMA,　when　incorporating　importance　sampling　and　distributed　computing,　evolved　into　our　method,　Advanced-COMA,　which　demonstrated　effectiveness　in　the　experiments.　Additionally,　a　fairness-aware　reward　function　was　designed　to　better　address　fairness　considerations.　Finally,　extensive　experiments　were　conducted　on　both　synthetic　and　real-world　traffic　networks,　including　grid　and　arterial　scenarios.　The　results　confirmed　the　superiority　of　our　methods　in　enhancing　traffic　conditions　and　fairness.　Some　details　can　be　found　on　https://github.com/caisibin/FS.　　©　2025　IEEE.