search

crowdAI / marLo

Multi Agent Reinforcement Learning using MalmÖ
MIT License
245 stars 46 forks source link

Standardise the action space building #30

Open spMohanty opened 6 years ago

spMohanty commented 6 years ago

Right now, the way the action space in an environment is built is pretty messy.

The ideal solution would be to delegate the action space building completely to the mission_spec, and have a separate helper function which takes a mission_spec as input and returns all the valid actions.

These actions can be returns as string "commands", and then we simply shape them as gym friendly classes (Discrete, etc etc).

cc. @katja-hofmann @AndKram

spMohanty commented 6 years ago

Related to https://github.com/crowdAI/marLo/issues/24

spMohanty commented 6 years ago

Addressed in https://github.com/crowdAI/marLo/commit/b3fe782943614aa26d78b4285239651cf288a42b Thanks @AndKram Will keep the issue open until this has been tested a bit more thoroughly.