Conference Paper This paper studies the Sensor Management (SM) problem for optical Space Object (SO) tracking. The tasking problem is formulated as a Markov Decision Process (MDP) and solved using Reinforcement Learning (RL). The RL problem is solved using the actor-critic policy gradient approach. The actor provides a policy which is random over actions and given by a parametric probability density function (pdf). The critic evaluates the policy by calculating the estimated total reward or the value function for the problem. The parameters of the policy action pdf are optimized using gradients with respect to the reward function. Both the critic and the actor are modeled using deep neural networks (multi-layer neural networks). The policy neural network takes the current state as input and outputs probabilities for each possible action. This policy is random, and can be evaluated by sampling random actions using the probabilities determined by the policy neural network's outputs. The critic approximates the total reward using a neural network. The estimated total reward is used to approximate the gradient of the policy network with respect to the network parameters. This approach is used to find the non-myopic optimal policy for tasking optical sensors to estimate SO orbits. The reward function is based on reducing the uncertainty for the overall catalog to below a user specified uncertainty threshold. This work uses a 30 km total position error for the uncertainty threshold. This work provides the RL method with a negative reward as long as any SO has a total position error above the uncertainty threshold. This penalizes policies that take longer to achieve the desired accuracy. A positive reward is provided when all SOs are below the catalog uncertainty threshold. An optimal policy is sought that takes actions to achieve the desired catalog uncertainty in minimum time. This work trains the policy in simulation by letting it task a single sensor to "learn" from its performance. The proposed approach for the SM problem is tested in simulation and good performance is found using the actor-critic policy gradient method.
展开▼