UNT Libraries - 2 Matching Results
Search Results
Note: All results matching your query require you to be a member of the UNT Community (you must be on campus or login with university credentials for access).
- Proximal Policy Optimization in StarCraft
- Deep reinforcement learning is an area of research that has blossomed tremendously in recent years and has shown remarkable potential in computer games. Real-time strategy game has become an important field of artificial intelligence in game for several years. This paper is about to introduce a kind of algorithm that used to train agents to fight against computer bots. Not only because games are excellent tools to test deep reinforcement learning algorithms for their valuable insight into how well an algorithm can perform in isolated environments without the real-life consequences, but also real-time strategy games are a very complex genre that challenges artificial intelligence agents in both short-term or long-term planning. In this paper, we introduce some history of deep learning and reinforcement learning. Then we combine them with StarCraft. PPO is the algorithm which have some of the benefits of trust region policy optimization (TRPO), but it is much simpler to implement, more general for environment, and have better sample complexity. The StarCraft environment: Blood War Application Programming Interface (BWAPI) is open source to test. The results show that PPO can work well in BWAPI and train units to defeat the opponents. The algorithm presented in the thesis is corroborated by experiments.
- Quantile Regression Deep Q-Networks for Multi-Agent System Control
- Training autonomous agents that are capable of performing their assigned job without fail is the ultimate goal of deep reinforcement learning. This thesis introduces a dueling Quantile Regression Deep Q-network, where the network learns the state value quantile function and advantage quantile function separately. With this network architecture the agent is able to learn to control simulated robots in the Gazebo simulator. Carefully crafted reward functions and state spaces must be designed for the agent to learn in complex non-stationary environments. When trained for only 100,000 timesteps, the agent is able reach asymptotic performance in environments with moving and stationary obstacles using only the data from the inertial measurement unit, LIDAR, and positional information. Through the use of transfer learning, the agents are also capable of formation control and flocking patterns. The performance of agents with frozen networks is improved through advice giving in Deep Q-networks by use of normalized Q-values and majority voting.