Abstract:
Reinforcement learning (RL) is the area of artificial intelligence research that has the goal of allowing autonomous agents to learn through interaction with an unknown environment. One of the core challenges in RL research is developing data efficient algorithms. In this talk I will describe recent work on increasing the data efficiency of Monte Carlo estimation in reinforcement learning. In RL, the data efficiency of Monte Carlo estimates is affected by how the learning agent takes actions and how the resulting data is weighted in the final approximation. I will first describe work on improving action selection. I will introduce two methods that adjust action selection based on past data to minimize the error in the final estimate. I’ll then discuss a re-weighting technique that allows RL agents to more efficiently use an already given set of samples. This technique can be seamlessly combined with the data collection methods from the first part of the talk to further minimize Monte Carlo estimation error in RL tasks.
Reinforcement learning (RL) is the area of artificial intelligence research that has the goal of allowing autonomous agents to learn through interaction with an unknown environment. One of the core challenges in RL research is developing data efficient algorithms. In this talk I will describe recent work on increasing the data efficiency of Monte Carlo estimation in reinforcement learning. In RL, the data efficiency of Monte Carlo estimates is affected by how the learning agent takes actions and how the resulting data is weighted in the final approximation. I will first describe work on improving action selection. I will introduce two methods that adjust action selection based on past data to minimize the error in the final estimate. I’ll then discuss a re-weighting technique that allows RL agents to more efficiently use an already given set of samples. This technique can be seamlessly combined with the data collection methods from the first part of the talk to further minimize Monte Carlo estimation error in RL tasks.