The objective of the proposed research is to develop non-collaborative spectrum sharing techniques that may be used by radars against dynamic and reactive emitters. This problem is important to military and commercial radio frequency (RF) applications, such as communications systems and radar, as demand for spectrum grows. Communications systems, particularly commercial cellular services, increasingly require higher bandwidth and more adaptive solutions to provide video streaming quality service to users across the world. As policy evolves, radars are being relegated to secondary user status in once dedicated spectrum bands and must have robust techniques to maximize radar performance while minimizing interference with primary users. It is important for policymakers to consider the degrees of freedom available across RF systems. Frequency, location, direction, time, polarization, and waveform code can be used to minimize interference and should be considered when defining spectrum access rules for secondary users. Many radars used for military applications are required to operate in a non-collaborative manner, because a collaborative approach to coexistence with commercial systems is at odds with the element of surprise inherent in many military radar applications. Existing methods to solve this problem involve sense and avoid techniques in the spectral domain, as well as some predictive methods that use statistical models to avoid interference while maximizing radar performance, usually through signal to interference plus noise ratio (SINR) and bandwidth joint optimization. The proposed research takes the concept of predictive spectrum sharing one step further by assuming that emitters may react to other spectrum users in their vicinity, leverages techniques to accelerate online learning convergence, and considers the benefits of modeling the problem as a Markov Decision Process (MDP) or further abstracting the model to include partial observability using a Partially Observable MDP (POMDP).