This work introduces SIPOMDPLite-net, a deep neural network (DNN) architecture for agentcontrol in partially observable multiagent settings with sparse interactions between agents. The
network represents a new method for planning in contexts modeled by the interactive partially
observable Markov decision process (I-POMDP) Lite and the decentralized sparse-interaction
MDP (Dec-SIMDP) frameworks, which facilitates self-interested planning in settings shared with
other agents more tractable than the well-known I-POMDP framework. The network uses fully-differentiable value iteration networks to simulate the solution of nested MDPs, which I-POMDP
Lite attributes to the other agent to model its behavior, avoiding the need for involving nondifferentiable techniques such as particle filtering to model the other agents more generally. We
train SIPOMDPLite-net on a small two-agent tiger-grid problem, for which it accurately learns
the underlying model and near-optimal policy, and the trained model continues to perform well on
much larger and complex grids. As such, SIPOMDPLite-net shows good transfer capabilities and
offers a lighter learning and planning approach for individual agents in multiagent settings.
