To be presented at 2023 European Conference on Artificial Intelligence (ECAI)
For supplementary material describing the baselines used in the paper and implementation details, please refer to the supplementary material at this link.
Following are some animations visualizing the behavior of CAST in comparison with the cost-agnostic baselines RSI and SPATS under different cost scenarios.
First, we show animations for a single trial of CAST, SPATS and RSI when the sensing cost dominates travel cost i.e. c_s = 50s for g=4 agents and k=3 targets in an 8x8 search space. For each agent, we visualize its posterior belief variance and the resulting action selected at the time step shown. The beta_hat shown at the top is the estimate based on all observations accumulated so far from all the agents. Total cost indicates in seconds the total sensing and travel cost incurred by all the agents.
CAST - Total cost in recovering all targets = 2479s
SPATS - Total cost in recovering all targets = 3635s
RSI - Total cost in recovering all targets = 2892s
Next, we show animations for a single trial of CAST, SPATS and RSI when the travel cost dominates sensing cost i.e. c_s = 0s for g=4 agents and k=3 targets in an 8x8 search space. For each agent, we visualize its posterior belief variance and the resulting action selected at the time step shown. The beta_hat shown at the top is the estimate based on all observations accumulated so far from all the agents. Total cost indicates in seconds the total sensing and travel cost incurred by all the agents.
CAST - Total cost in recovering all targets = 123s
SPATS - Total cost in recovering all targets = 449s
RSI - Total cost in recovering all targets = 308s