Here is a
problem data file with descriptions of some of the benchmark problems
we have used. The recycling robot problem (introduced in the UAI-07 paper by
Amato, Bernstein and Zilberstein) is case 9, the broadcast channel problem
(introduced in the IJCAI-05 paper by Bernstein, Hansen and Zilberstein) is
case 10, the two agent tiger problem is case 22 (introduced as Tiger-A in
Nair et al. IJCAI-03), the 3x3 grid problem is case 63 (from Amato,
Dibangoye and Zilberstein ICAPS-09), the stochastic Mars rover problem is
case 78 (from Amato and Zilberstein AAMAS-09) and the box pushing problem is
case 99 (from Seuken and Zilberstein UAI-07). The discount factor that we
have used in each problem is 0.9.

Nonlinear programming (NLP) formulations and problem data files

Moore NLP formulation for POMDPs - download
Moore NLP formulation for DEC-POMDPs - download
General Mealy NLP formulation for POMDPs (note that this is the full model
without removing variables and constrainsts based on exploiting problem
structure) - download
General Mealy NLP formulation for DEC-POMDPs (note that this is the full
model without removing variables and constrainsts based on exploiting
problem structure) - download

Problem Data

Note that these files are without initialization. Typically, we have used
random deterministic controllers which are then optimized. This can be set
in the problem data file. Also, the number of nodes can be changed on the
first line of the file.

The table below provides the highest known values for a range of
benchmark problems with the common discount factor of 0.9 (except Wireless Network
which used 0.99). We also list the paper that each result first appeared.
Note that the results are often an average over a number of runs, so
single runs may have higher values than those listed here. Optimal values
for finite horizon problems can be found here.

Mark Gruman has developed several tools for the creation and manipulation of
multiagent domains. The format
is based on the input files found on
Tony's POMDP Page .