Research On Applying Deep Reinforcement Learning In Image Based Control And Image Classification Tasks

In recent years,new advances in deep learning and reinforcement learning have addressed several challenges in combining these two techniques and brought deep reinforcement learning great success.Deep reinforcement learning algorithms have both strong generalization and feature extraction abilities and the ability to bootstrap task-solving policies.Recent work shows that deep reinforcement learning agents can even outperform humans in several control tasks.However,limitations still exist.In terms of the depth perspective,existing algorithms generally fail in learning multiple tasks or compound tasks in a continuous action space.Also,in terms of the breadth perspective,little work has been done in applying deep reinforcement learning in computer vision area,especially in image classification tasks.Work in this thesis is aimed to research on both these two aspects.Three algorithms have been proposed in this thesis: multi-DDPG,h-DDPG and selfreinforced network(SRN)algorithms.The first two are for image based multi-task learning,hierarchical learning in a continuous action space and the last one is for image classification.The multi-DDPG algorithm comprises a single-critic,multi-actor architecture,with each actor being responsible for learning one task.This architecture has been further extended in h-DDPG algorithm to a dual-critic,multi-actor architecture,adding a new critic to learn compound task by combining different actors.The SRN algorithm is an attempt to apply deep reinforcement learning in image classification tasks.In SRN algorithm,an agent is responsible to decide the time that an input image will be classified.If the agent decides that the image cannot be classified yet,a transformation will be chosen and applied on the image before a second round classification by the classifier network.Experiments have been done to test all three algorithms.For multi-DDPG,results show that it can control the agent to learn multiple task policies that are comparable to the ones learned by single task learning algorithm,while this performance stays consistent when the number of task or the number of the constraints in the tasks increase.For h-DDPG,results show that it can effectively learn multiple basic tasks and a compound task simultaneously.The policies learned for solving the compound task are even better than discrete action based algorithms.For SRN,results show that the existing of the decision agent can reduce the error rate of the classifier network by 18.82%.