Dynamic malicious jamming of spectrum throughout the entire communication band is a major issue confronted by tactical communications. The conventional spectrum access scheme based on the central control node fails to satisfy the demand of tactical communications; it has a number of drawbacks including low battlefield survival rates, high computational complexity, large interactive communication overhead and slow reaction to dynamic jamming. In this paper, we propose a distributed multi-agent spectrum access strategy without the interactive communication overhead to evade dynamic jamming. Furthermore, A simplified Q reinforcement learning algorithm is applied to alleviate collisions of spectrum among nodes. Under the proposed strategy, all nodes can predict and evade dynamic jamming as well as avoid collisions of spectrum usage with other nodes. Simulation results verify the collision-avoidance learning algorithm and indicate that the proposed strategy outperforms the random spectrum access strategy.