1. OS: Ubuntu
2. CPU: Intel i7
3. RAM: 48 GB
4. GPU: Nvidia TitanX
5. Cores: 12
6. Implementation: Python, Jupyter Notebook
7. DLR Agents Implementation: Rllab
8. LSTM Implementation: Keras
9. EFC, HLR, GPL Implementation: OpenAI Gym
10. Statistical Analysis: Python and R

San Diego State University

1. Dataset
2. System and Tools
3. Experiments

Method and Experiments (Using deep reinforcement learning for personalizing review sessions on e-learning platforms with spaced repetition)

In this research synthetic dataset was used. The simulators: Half-Life Regression (HLR), GPL (General-Power Law) , Exponential Forgetting Curve (EFC) were used to generate the data. The data had following representation: {q, a, t, d}:

1. q - most recent item shown to the student
2. a - [0, 1]  incorrect/correct answer
3. t - timestamp
4. d - delay between most recent item and before (Set to 5s)

For LSTM the data was generated by EFC. The data was generated in 3 different ways:

1. Random Sample: exercises and answers by students were randomly collected
2. Random Policy Tutor: the exercises were chosen by the policy at random while student answer weren't 
3. Supermemo Tutor: Supermemo algorithm was used


Dataset (Using deep reinforcement learning for personalizing review sessions on e-learning platforms with spaced repetition)

Systems and Tools (Using deep reinforcement learning for personalizing review sessions on e-learning platforms with spaced repetition)

1. Experimental Setup
2. Reward functions and performance metrics 
3. Training the LSTM
4. Relation between rewards and thresholds
5. Performance of RL agent when the number of items are varied
6. Performance of TRPO vs. TNPG algorithms
7. Comparison between likelihood and average of sum of outcomes based reward functions (research objective)
8.  Performance of TRPO with reward shaping (research objective)
9. Evaluation 

Learn Before

Related