(Georgia Institute of Technology, 2020-12)
Gasser, Tarik
Training an Artificial Intelligence could be challenging in so many ways. In our research we are building a strong AI that has the ability to make decision on its own without given explicit answers. We are using Reinforcement learning with Imitating learning to train the Artificial Intelligent offline before putting him through the environment.
(Georgia Institute of Technology, 2019-05)
Tejani, Farhan
We present a differentiable framework capable of learning a wide variety of compositions of simple policies that we call skills. By recursively composing skills with themselves, we can create hierarchies that display complex behavior. Skill networks are trained to generate skill-state embeddings that are provided as inputs to a trainable composition function, which in turn outputs a policy for the overall task. Our experiments on an environment consisting of multiple collect and evade tasks show that this architecture is able to quickly build complex skills from simpler ones. Furthermore, the learned composition function displays some transfer to unseen combinations of skills, allowing for zero-shot generalizations. We also design experiments for learning both skills as well as optimal compositions of those skills from scratch. This allows for the agent to more intelligently search the space of possible skills rather than having a human hand-design the composition functions, which may be limited in scope.
(Georgia Institute of Technology, 2018-05)
Kumar, Saurabh
Typical reinforcement learning (RL) agents learn to complete tasks specified by reward functions tailored to their domain. As such, the policies they learn do not generalize even to similar domains. To address this issue, we develop a framework through which a deep RL agent learns to generalize policies from smaller, simpler domains to more complex ones using a recurrent attention mechanism. The task is presented to the agent as an image and an instruction specifying the goal. This meta-controller guides the agent towards its goal by designing a sequence of smaller subtasks on the part of the state space within the attention, effectively decomposing it. As a baseline, we consider a setup without attention as well. Our experiments show that the meta-controller learns to create sub-goals within the attention. These results have implications for human-robot interactive applications, in which a robot can transfer skills it has learned in one task to another one and be robust to variability in its environment.
(Georgia Institute of Technology, 2017-08)
Sood, Srijan
Reinforcement learning is a powerful mechanism for training artificial and real-world agents to perform tasks. Typically, one can define a task for an agent by simply specifying rewards that reflect the agent’s performance. However, each time the task changes, one must develop a new reward specification. Our work aims to remove the necessity of designing rewards in tasks consisting of visual inputs. When humans are learning to complete tasks, we often look to other sources for inspiration or instruction. Even if the representation is different from our own, we can adapt our own representation to the task representation. This motivates our own work, where we present tasks to an agent that are from an environment different than its own. We compare the cross-domain goal representation with the agents representation to form Cross-Domain Perceptual Reward (CDPR) functions and show that these enable the agent to successfully complete its task.