It　is　anticipated　that　roadside　infrastructure　sensing　units　(ISU)　can　cooperatively　work　with　intelligent　and　connected　vehicles　(ICVs)　to　perceive　traffic　scenes　more　accurately　when　ICVs　increasingly　penetrate　the　market.　However,　the　dynamic　occlusion　issue　may　still　impair　the　cooperative　vehicle-infrastructure　perception　capability　(CVIPC).　This　study　addresses　the　lack　of　an　effective　method　for　placing　ISUs　to　augment　CVIPC　in　a　partially　connected　traffic　environment.　This　study　introduces　probabilistic　occupancy　grids　(POGs)　to　model　the　uncertainty　of　dynamic　occlusions.　The　ground　truth　POG　is　estimated　with　a　co-simulation　method,　while　the　observed　POG　by　ISUs　and　ICVs　are　estimated　using　the　proposed　occlusion-considered　ray-tracing　algorithm.　The　cross　entropy　(CE)　is　applied　to　measure　the　difference　between　the　ground　truth　and　observed　POGs　and　is　used　as　a　surrogate　metric　for　estimating　CVIPC.　Setting　ISUs’　placement　parameters　and　POG-based　CE　as　decision　variables　and　the　objective,　respectively,　Bayesian　optimization　(BO)　is　integrated　with　the　multi-agent　deep　reinforcement　learning　(DRL)　to　maximize　CVIPC.　The　test　results　imply　that　combining　BO　and　DRL　can　outperform　BO　in　optimizing　ISUs’　placement.　Compared　to　the　simulation-in-the-loop　optimization,　the　surrogate　metric　-based　framework　can　achieve　faster　optimization　with　a　small　compromise　on　the　optimized　CVIPC　measured　by　intersection　over　union.　Traffic　volume,　traffic　composition　and　ICV　penetration　rate　all　substantially　affect　CVIPC.　In　the　test　cases,　as　the　ICV　penetration　rate　reaches　50%,　the　observed　POG　is　very　close　to　the　ground　truth　POG,　and　a　further　increase　in　the　number　of　ICVs　does　not　substantially　contribute　to　improving　CVIPC.　©　2014　IEEE.