drumpt / AI-Soccer-Code-Generator

Apache License 2.0

1 stars 1 forks source link

To Do - Deep Learning System #5

Open lfelipesv opened 4 years ago

lfelipesv commented 4 years ago

To Do (Urgent)

[x] 3 types of State functions - Code Template
[x] 3 types of Action functions - Code Template
[x] 3 types of Reward functions - Code Template
[x] Finish code template for single-agent DDPG algorithm
[x] Finish code template for multi-agent QMIX algorithm
[x] Solve CSS and div problems for the hyperparameter boxes
[x] Finish code generation for current DQN and IQL code templates

For the NEAR future

[ ] Create a flexible reward framework: the user can design the reward function using blocks or an input in the website to design mathematical functions

Important Design Question (to discuss)

The deep learning system is an incipient type of system (to teach and implement deep learning algorithms). The main objectives of this system are to:

model reward function easily
choose RL algorithm and hyperparameters easily

How to expand the algorithm 'block' to facilitate the understanding by the user? For example, how Chris Olah's explain the LSTM network...

For the future

[ ] Add more RL algorithms
[ ] Change how to choose state, action and reward

lfelipesv commented 4 years ago

I added the code generator for DQN and IQL.

Still dummy:

State, Action, Reward choice
Robots to train in multi-agent system
Choosing algorithm block