Market-Based Allocation with Indivisible Bids - Agent-Mediated Electronic Commerce - page 64

• PC f , r : Production capacity per part type, the quantity of parts of type r that the

module produces per day.

•

RC f : Reconfiguration cost to be paid if the configuration is used.

• RT f : Reconfiguration time (in days) that it takes to set up configuration f , dur-

ing which no part can be produced.

The configuration capacities and times, along with the period length L ,definethe
production possibilities for module i . The various cost parameters define the total cost
for any feasible production plan.
Although complicated, the foregoing determines well-defined optimization problems
for the agent:

- Determining an optimal production plan given holdings of goods r .
- Determining optimal demand for goods r given current holdings and market prices.

3.3 Market Configuration

The overall market system comprises the agents representing manufacturing modules,
plus one auction for each part type. We simulate an instance of this setup by gener-
ating parameter values from prespecified probability distributions, and communicating
these values to the respective agents. Each agent is initially allocated customer orders
corresponding to equal shares, D r / N , of the overall demand for each part r .
The simulations are implemented using our configurable market game server, AB3D
[13]. Each game instance lasts twenty minutes, with each auction clearing periodically
every 48 seconds. The auctions are staggered, so that the initial clears occur at multiples
of 48 / Mseconds .
The agents operate asynchronously, submitting bids to the auctions according to the
policy described in Section 4.2. Agents can request price quotes reflecting the latest
auction state, and retrieve notices of any transactions from prior bids.
At the end of a game instance, the server calculates final holdings based on cumula-
tive transactions, and determines a score for each agent. The score depends on an agent's
production plan given its total available orders, which entails solving an optimization
problem for each agent. AB3D solves these using a commercial optimization package
(AMPL/CPLEX), given an integer linear programming (ILP) formulation specified as
part of the game description.
The overall value of the resulting allocation is simply the sum of the scores over
the N agents. For comparison, we can also calculate (offline if necessary) the global
optimum of the system without trading, assuming a central planner that can allocate
orders across manufacturing modules.

4 Experiments

We ran a set of 58 paired trials with both standard and AON auctions. For AON auctions,
we tested standard , marginal ,and full schedule quotes. The following sections describe
the specific problem instance we chose for our manufacturing scenario, the behavior of
the agents, and the results obtained.