Using RL, automate many battlefield decisions

• Key concepts: agent, environment, state, observation, action, reward, policy, Q-value.
• Extensions: multi-agent RL, hierarchical RL, partially observed RL.

• Hierarchy over actions (combine low level actions into high level primititive) – domain specific
• Intrinsic rewards for training (solving one task may not transfer to others)
• Our application?

Learns a span of skills in pretraining environment, then (later) higher level network switches between these skills.

• Our application?

• Architecture (concatenation versus bilinear integration)

• Additive reward bonus to regularize and keep skills distinct.
• Our application?

• Frozen low-level policies (not necessarily so)