Human driven vehicles (HDVs) with selfish objectives cause low traffic efficiency in an un-signalized intersection. On the other hand, autonomous vehicles can overcome this inefficiency through perfect coordination. In this paper, we propose an intermediate solution, where we use vehicular communication and a small number of autonomous vehicles to improve the transportation system efficiency in such intersections. In our solution, two connected autonomous vehicles (CAVs) lead multiple HDVs in a double-lane intersection in order to avoid congestion in front of the intersection. The CAVs are able to communicate and coordinate their behavior, which is controlled by a deep reinforcement learning (DRL) agent. We design an altruistic reward function which enables CAVs to adjust their velocities flexibly in order to avoid queuing in front of the intersection. The proximal policy optimization (PPO) algorithm is applied to train the policy and the generalized advantage estimation (GAE) is used to estimate state values. Training results show that two CAVs are able to achieve significantly better traffic efficiency compared to similar scenarios without and with one altruistic autonomous vehicle.