Research on Self-Gaming Training Method of Wargame Based on Deep Reinforcement Learning

In this paper, we propose a novel conditional-generativeadversarial-nets-based image captioning framework as an extension of traditional reinforcement-learning (RL)-based encoder-decoder architecture. To deal with the inconsistent evaluation problem among different objective language metrics, we are motivated to design some “discriminator” networks to automatically and progressively determine whether generated caption is human described or machine generated. Two kinds of discriminator architectures (CNN and RNNbased structures) are introduced since each has its own advantages. The proposed algorithm is generic so that it can enhance any existing RL-based image captioning framework and we show that the conventional RL training method is just a special case of our approach. Empirically, we show consistent improvements over all language evaluation metrics for different state-of-the-art image captioning models. In addition, the well-trained discriminators can also be viewed as objective image captioning evaluators.

Download Full-text

Two-stage population based training method for deep reinforcement learning

Proceedings of the 3rd International Conference on High Performance Compilation, Computing and Communications - HP3C '19 ◽

10.1145/3318265.3318294 ◽

2019 ◽

Author(s):

Yinda Zhou ◽

Weiming Liu ◽

Bin Li

Keyword(s):

Reinforcement Learning ◽

Population Based ◽

Training Method ◽

Two Stage

Download Full-text

Spellcaster Control Agent in StarCraft II Using Deep Reinforcement Learning

Electronics ◽

10.3390/electronics9060996 ◽

2020 ◽

Vol 9 (6) ◽

pp. 996

Author(s):

Wooseok Song ◽

Woong Hyun Suh ◽

Chang Wook Ahn

Keyword(s):

Reinforcement Learning ◽

Force Field ◽

Real Time ◽

Transfer Learning ◽

Main Idea ◽

Control Agent ◽

Training Methods ◽

Training Method ◽

Combat Training

This paper proposes a DRL -based training method for spellcaster units in StarCraft II, one of the most representative Real-Time Strategy (RTS) games. During combat situations in StarCraft II, micro-controlling various combat units is crucial in order to win the game. Among many other combat units, the spellcaster unit is one of the most significant components that greatly influences the combat results. Despite the importance of the spellcaster units in combat, training methods to carefully control spellcasters have not been thoroughly considered in related studies due to the complexity. Therefore, we suggest a training method for spellcaster units in StarCraft II by using the A3C algorithm. The main idea is to train two Protoss spellcaster units under three newly designed minigames, each representing a unique spell usage scenario, to use ‘Force Field’ and ‘Psionic Storm’ effectively. As a result, the trained agents show winning rates of more than 85% in each scenario. We present a new training method for spellcaster units that releases the limitation of StarCraft II AI research. We expect that our training method can be used for training other advanced and tactical units by applying transfer learning in more complex minigame scenarios or full game maps.

Download Full-text