README.md 1.28 KB
Newer Older
hcy20's avatar
hcy20 committed
1
2
# DRL Risk

Hudson Yeo's avatar
Hudson Yeo committed
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Repository for the individual project: Distributional Reinforcement Learning and the meaning of uncertainty in the return distribution.
In the folder code, there are several Jupyter notebooks that contain the full code for various aspects in the project. Below, we describe the main files used in the report.

| File      | Description |
| ----------- | ----------- |
| c51      | Main c51 file for WCW env       |
| mdp  | c51 agent for MDP env in Risk-aware action selection        |
| safety_classifier_investigation | All code used for investigating various SCs in report, except QSC |
| DQN | QSC investigations |
| blackjack | SC investigations on Blackjack |
| Risky_MDP | Risky MDP env with c51 agent |
| polgrad_risky_mdp | Multi-Agent SC on Risky MDP env |
| polgrad_wcw | Multi-Agent SC on WCW env, including its variants such as Inverted WCW and MWCW |
| IQN | IQN code, inclusive of monotonically increasing quantile function |
Hudson Yeo's avatar
Hudson Yeo committed
17
18
19
20
21
| QRDQN | QRDQN code, for graphs |

The only libraries required for this project are numpy, torch, and gym. Additional libraries are not used.
All code in this repository was written to be used on Google Colab, which runs Python 3.7.11 as checked on 2 September 2021.
The code was also tested on Computing lab machines, such as gpu12, and they run fine.