Multi-agent　simulation　based　on　deep　reinforcement　learning　can　simulate　realistic　economic　agents　well,　and　can　play　an　important　role　in　research　work　on　complex　economic　system.　In　the　simulated　economic　system　model,　DQN　is　used　to　handle　the　decision-making　of　economic　agents,　and　their　behavior　strategies　are　continuously　improved　with　the　accumulation　of　experience　as　they　interact　with　the　environment.　However,　the　action　selector　used　by　economic　agents　can　affect　their　policy　changes.　Taking　this　problem　into　consideration,　this　paper　did　research　on　the　action　selector　used　by　economic　agents.　Through　simulation　experiments,　we　compare　the　behavior　policy　of　economic　agents　using　different　action　selectors　including　the　traditional　-greedy　and　a　novel　NoisyNet　method.　The　results　show　that　an　appropriate　action　selector　enables　economic　agents　to　explore　more　widely　and　comprehensively,　and　learn　a　better　policy.　©　2021　IEEE.