Train agents in various environment using unity ml-agents.
Each environment is implemented as a submodule and can be found in the sections below.
Policies, hyperparameters, and scripts are described in the blog post.
Env | Detail Description | Submodule Repo |
---|---|---|
Knight Battle | mlagent_09 | ml-playground-battle |
Domination | mlagent_08 | - |
Predator and Prey | mlagent_07 | - |
Obstacle Avoidance | mlagent_01 ~ 06 | - |
- Train agent to do melee combat (1 vs 1, many vs many) in limited area.
- Learning with MA-POCA, self-play (co-op & competitive)
- Agents can use sword to take attack or use shield to block attack.
- Also, you can play with trained agents. Beat them out!
- Repository : ml-playground-battle
- Detail Description : Blog Post
- Train agent to fill tiles with its color for win.
- Learning with MA-POCA, self-play
- Each time the agent moves, it paints a tile with its color.
- When the closed path is complete, all of the interior tiles are painted with its color.
- Detail Description : Blog Post
- Train Predator to catch Prey, Train Prey to escape from Predator.
- Predator wins if it catches Prey, and Prey wins if it alive at the end of time.
- Learning with MA-POCA (Multi-Agent Posthumous Credit Assignment), self-play
- Detail Description : Blog Post
- Train agent to avoid obstacles and reach the goal in various environments.
- The preview above is a snippet, the full list is at here