Meima, Niels and Mallon, Sjors (2018) Effectiveness of Connectionist Q-learning Strategies on Agent Performance in Asteroids. Bachelor's Thesis, Artificial Intelligence.
|
Text
AI_BA_2018_NIELSMEIMA.pdf Download (604kB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (98kB) |
Abstract
This research proposes a higher-order state extraction algorithm serving as input for neural networks to learn to play the Atari game Asteroids. Asteroids is a 1980's space shooter, and poses a challenging environment due to its continuous and stochastic nature. Learning capabilities of the reinforcement learning algorithms Q-learning, Q-learning combined with a target network, Double Q-learning, QV-learning and QVMAX-learning are compared at a constant difficulty level, both using online learning and experience replay. Q-learning combined with a target network achieved the highest win rate of 0.76, in both the online and experience replay setting. Furthermore, the influence of incremental learning on agent performance is compared to learning at a constant difficulty. Incremental learning did not show a significant improvement in performance. Finally, state modeling in combination with Monte Carlo rollouts is used to learn from predictions about the future. Results show that learning from predictions is ineffective in its current implementation. The agent effectively learns to play the game Asteroids using the higher-order state extraction algorithm in combination with the described reinforcement learning algorithms.
| Item Type: | Thesis (Bachelor's Thesis) |
|---|---|
| Supervisor name: | Wiering, M.A. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Bachelor's Thesis |
| Language: | English |
| Date Deposited: | 28 Jul 2018 |
| Last Modified: | 30 Jul 2018 14:08 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/18114 |
Actions (login required)
 |
View Item |