Learning to Generalise in Sparse Reward Navigation Environments

Asad Jeewa, Anban W. Pillay, Edgar Jembere

Abstract: It is customary for RL agents to use the same environments for both training and testing. This causes the agents to learn specialist policies that fail to generalise even when small changes are made to the training environment. The generalisation problem is further compounded in sparse reward environments. This work evaluates the efficacy of curriculum learning for improving generalisation in sparse reward navigation environments: we present a manually designed training curriculum and use it to train agents to navigate past obstacles to distant targets, across several hand-crafted maze environments. The curriculum is evaluated against curiosity-driven exploration and a hybrid of the two algorithms, in terms of both training and testing performance. Using the curriculum resulted in better generalisation: agents were able to find targets in more testing environments, including some with completely new environment characteristics. It also resulted in decreased training times and eliminated the need for any reward shaping. Combining the two approaches did not provide any meaningful benefits and resulted in inferior policy generalisation.