Efficient dynamic spectrum access is important for next-generation wireless networks. This paper develops a decentralized multi-agent deep enforcement learning approach for cognitive radios to opportunistically access the spectrum. We formulate the problem as a stochastic game between distributed learning agents. The agents operate independently to optimize channel selection and power allocation policies using deep Q-learning. We evaluate our algorithm in real-world wireless environments with time-varying channel conditions and traffic patterns. Results illustrate that the proposed scheme significantly improves spectrum utilization compared to traditional sensing-based approaches. Our work demonstrates the promise of multi-agent reinforcement learning for distributed dynamic spectrum access.