Akakzia, Mr Ahmed (2019) Goal-Conditioned Meta-Reinforcement Learning with Replay PFE - Project Graduation, ENSTA.



Artificial Intelligence gave birth to a revolution, stimulated by better and more efficient machine learning algorithms. However, in contrast to Humans, training neural networks requires a huge quantity of computations. Moore’s law highlights that hardware capabilities would grow tremendously, leading to better computational strength. But still, this is not really intuitive and is not really correlated with the Humans learning abilities that allow them to acquire new concepts from only a few examples and quickly adapt to previously unseen situations. Hence, in Artificial Intelligence, dealing with multiple tasks is a challenge to better converge to Humans’ behaviour. When tasks and goals are not known in advance, an agent may use either multitask learning or meta reinforcement learning to learn how to transfer knowledge from what it learned before. Recently, goal-conditioned policies and hindsight experience replay have become standard tools to address transfer between goals in the multitask learning setting. In this report, we show that these tools can also be imported into the meta reinforcement learning when one wants to address transfer between tasks and between goals at the same time. More importantly, we investigate whether the meta reinforcement learning approach brings any benefit with respect to multitask learning in this specific context. Our study based on a basic meta reinforcement learning framework reveals that showcasing such benefits is not straightforward, calling for further comparisons between more advanced frameworks and richer sets of tasks.

Item Type:Thesis (PFE - Project Graduation)
Subjects:Mathematics and Applications
ID Code:7956
Deposited By:Ahmed Akakzia
Deposited On:09 nov. 2020 15:24
Dernière modification:09 nov. 2020 15:24

Repository Staff Only: item control page