We focus on the task transfer in reinforcement learning and specifically in Q-learning. There are three main model free methods for performing task transfer in Q-learning: direct transfer, soft transfer and memory-guided exploration. In direct transfer, the Q-values from a previous task are used to initialize the Q-values of the next task. The soft transfer initializes the Q-values of the new task with a weighted average of the standard initialization value and the Q-values of the previous task. In memory-guided exploration the Q-values of previous tasks are used as a guide in the initial exploration of the agent. The weight that the agent gives to its past experience decreases over time. We explore stability issues related to the off-policy nature of memory-guided exploration and compare memory-guided exploration to soft transfer and direct transfer in three different environments.
展开▼