MuKumari

This is intended as an easy-to-use playground for quickly setting up multi-agent scenarios in simulated environments with user-specifiable characteristics.

TODOs:

Steps necessary to get Crux's IQLearn working:

• 1. Fix the state/observation vectorization conversion via the POMDPs convert_s function.
  • Proposition 1.1: Change the MDP type into a POMDP and use the initialobs to vectorize the state into some observations.
  • Proposition 1.2: Remove the observation output from the gen, making KAgentMDP a true MDP; self-define convert_s to vectorize the state.
• 2. Decide what to do w.r.t. the ContinuousSpace converter via Crux's state_space function
  • Proposition 2.1: Leave the default values for μ and σ, in hopes that the statespace will be refined by the network over time
  • Proposition 2.2: Edit the values to somehow better match the world (ex. make μ=5. and σ=5., etc.)
• 3. Generate "expert trajectories", wrapping simulator output in an Crux's ExperienceBuffer type.