The Radar Resource Management (RRM) problem in a multi-sensor multi-target scenario is considered. The problem is defined as a constrained optimization problem in which the predicted error covariance is minimized subject to resource budget constraints. By applying Lagrangian Relaxation (LR) the problem is decoupled into multiple sub-optimization problems. The problem is modeled according to a Partially Observable Markov Decision Process (POMDP). Using a stochastic optimization framework called policy rollout, the POMDP is solved non-myopically by looking ahead into the expected future. Two novel implementations, namely a centralized and distributed implementation are presented as viable approaches for solving this problem for a multi-sensor case. The centralized implementation, defined as the approximately optimal solution, utilizes a global policy per task. As such, the policy rollout for a single target needs to explore the actions of multiple sensors. The distributed implementation is considered as a practical alternative to improve on the computational complexity of the policy rollout of the centralized implementation. Now per sensor and per task a policy rollout is computed. To maintain a similar performance as the centralized implementation, at the beginning of each policy rollout the last known actions of the other sensors are shared. An additional third independent implementation is considered. The independent implementation uses no communication during the optimization process and is considered to be the implementation with the lowest performance with respect to the cost. All implementations have been applied to multiple two-dimensional simulated radar tracking scenarios. A comparison is made between the centralized, distributed and independent implementation based on the average cost and runtime. Results indicate that both the centralized and distributed implementation outperform the independent implementation with respect to the cost by a factor two. Subsequently, the distributed solution converges to similar results as the centralized implementation while requiring significantly less computational resources.