TODOs and Ideas

Todos

• By Aug 10th

  • publish results (play video, blogpost)

• By Aug 3rd:

• By July 27th:

• By July 20th: 9:30pm:

  • DH: modularize paac.py
  • DH: integrate CTS to paac.py
  • [?] Sangjin: try MOL bonus
  • [?] HG: try feature control bonus

• July 5th Wed, 6:10-7:10pm:

  • play Montezuma's Revenge yourself
  • train any algorithm of your choice on Montezuma's Revenge and share results (testing environment shared here)
  • (optional) read NLP-based solution to MZ
  • (optional) read why MZ is hard

• By July 15th Sat, 2-4pm:

  • Hoyeop Kim: Count-Based Exploration 2017
  • Sangjin Park: Hierarchical RL
  • DH: FeUdal Networks
  • HG: [UNREAL](https://arxiv.org/abs/1611.05397]

ideas

• reward function learning (inverse RL)
• bonus for surviving
• objects detection
• attention model
• transfer learning
• life -> give larger penalty for losing a life
• semantic segmentation (edge) difference as bonus