Vlad Ilie
Hard-working, ambitious and enthusiastic student, capable of quick learning. Having solid time management, strong organizational and multitasking skills, I am dedicated to completing tasks on time.At the moment I am working on my MSc Thesis: implementing a Deep Reinforcement Learning agent in StarCraft II
less
Related Authors
Interests
Uploads
Papers by Vlad Ilie
To overcome these challenges, the project uses the Reaver framework to set up an initial baseline Synchronous Advantage Actor-Critic (A2C) agent, which is then modified to use the RoE method. The project uses the game's non-spatial features (NSF), including the game score, to define an intrinsic reward function associated with the start, progress and completion of actions. Having the algorithm score its own ability to play the game based on environmental changes allows the agent to perform an automated curriculum learning process, where it attempts gradually more complex tasks.
The RoE method has already been tested and proven successful for the VizDoom environment, this project extends its functionality to SC2 and implements three variations of it: Binary RoE, Quantitative RoE and Greedy RoE. Combinations of the three options can be used on the same agent and can be tailored for specific purposes, so the agent is guided towards desired behaviours.
To overcome these challenges, the project uses the Reaver framework to set up an initial baseline Synchronous Advantage Actor-Critic (A2C) agent, which is then modified to use the RoE method. The project uses the game's non-spatial features (NSF), including the game score, to define an intrinsic reward function associated with the start, progress and completion of actions. Having the algorithm score its own ability to play the game based on environmental changes allows the agent to perform an automated curriculum learning process, where it attempts gradually more complex tasks.
The RoE method has already been tested and proven successful for the VizDoom environment, this project extends its functionality to SC2 and implements three variations of it: Binary RoE, Quantitative RoE and Greedy RoE. Combinations of the three options can be used on the same agent and can be tailored for specific purposes, so the agent is guided towards desired behaviours.