next up previous
Next: Discussion Up: Controlled Experiments Previous: Varying the Predictor

ATTac-2001 vs. EarlyBidder

In a sense, the two agents that finished at the top of the standings in TAC-01 represented opposite ends of a spectrum. The livingagents agent uses a simple open-loop strategy, committing to a set of desired goods right at the beginning of the game, while ATTac-2001 uses a closed-loop, adaptive strategy.

The open-loop strategy relies on the other agents to stabilize the economy and create consistent final prices. In particular, if all eight agents are open loop and place very high bids for the goods they want, many of the prices will skyrocket, evaporating any potential profit. Thus, a set of open-loop agents would tend to get negative scores--the open-loop strategy is a parasite, in a manner of speaking. Table 13 shows the results of running 27 games with 7 copies of the open-loop EarlyBidder and one of ATTac-2001. Although motivated by livingagents, in actuality it is identical to ATTac-2001 except that it uses $\mbox{\sf {SimpleMean}}_{ev}$ and it places all of its flight and hotel bids immediately after the first flight quotes. It bids only for the hotels that appear in $G^*$ at that time. All hotel bids are for $1001. In the experiments, one copy of $\mbox{\sf {ATTac-2001}}_{s}$ is included for comparison. The price predictors are all from Phase I in the preceding experiments. EarlyBidder's high bidding strategy backfires and it ends up overpaying significantly for its goods. As our experiments above indicate, ATTac-2001 may improve even further if it is allowed to train on the games of the on-going experiment as well.

Table 13: The results of running ATTac-2001 against 7 copies of EarlyBidder over the course of 27 games. EarlyBidder achieves high utility, but overpays significantly, resulting in low scores.
Agent Score Utility
ATTac-2001 2431 $\pm $ 464 8909 $\pm $ 264
EarlyBidder(7) $-4880$ $\pm $ 337 9870 $\pm $ 34

The open-loop strategy has the advantage of buying a minimal set of goods. That is, it never buys more than it can use. On the other hand, it is susceptible to unexpected prices in that it can get stuck paying arbitrarily high prices for the hotel rooms it has decided to buy.

Notice in Table 13 that the average utility of the EarlyBidder's clients is significantly greater than that of ATTac-2001's clients. Thus, the difference in score is accounted for entirely by the cost of the goods. EarlyBidder ends up paying exorbitant prices, while ATTac-2001 generally steers clear of the more expensive hotels. Its clients' utility suffers, but the cost-savings are well worth it.

Compared to the open-loop strategy, ATTac-2001's strategy is relatively stable against itself. Its main drawback is that as it changes its decision about what goods it wants and as it may also buy goods to hedge against possible price changes, it can end up getting stuck paying for some goods that are ultimately useless to any of its clients.

Table 14 shows the results of 7 copies of ATTac-2001 playing against each other and one copy of the EarlyBidder. Again, training is from the seeding round and finals of TAC-01: the agents don't adapt during the experiment. Included in this experiment are three variants of ATTac-2001, each with a different flight-lookahead parameter (from the section on ``cost of postponing flight commitments''). There were three copies each of the agents with flight-lookahead set to 2 and 3 (ATTac-2001(2) and ATTac-2001(3), respectively), and one ATTac-2001 agent with flight-lookahead set to 4 (ATTac-2001(4)).

Table 14: The results of running the EarlyBidder against 7 copies of ATTac-2001 over the course of 197 games. The three different versions of ATTac-2001 had slightly different flight-lookaheads.
Agent Score Utility
EarlyBidder 2869 $\pm $ 69 10079 $\pm $ 55
ATTac-2001(2) 2614 $\pm $ 38 9671 $\pm $ 32
ATTac-2001(3) 2570 $\pm $ 39 9641 $\pm $ 32
ATTac-2001(4) 2494 $\pm $ 68 9613 $\pm $ 55

From the results in Table 14 it is clear that ATTac-2001 does better when committing to its flight purchases later in the game (ATTac-2001(2) as opposed to ATTac-2001(4)). In comparison with Table 13, the economy represented here does significantly better overall. That is, having many copies of ATTac-2001 in the economy does not cause them to suffer. However, in this economy, EarlyBidder is able to invade. It gets a significantly higher utility for its clients and only pays slightly more than the ATTac-2001 agents (as computed by utility minus score).9

The results in this section suggest that the variance of the closing prices is the largest determining factor between the effectiveness of the two strategies (assuming nobody else is using the open-loop strategy). We speculate that with large price variances, the closed-loop strategy (ATTac-2001) should do better, but with small price variances, the open-loop strategy could do better.

next up previous
Next: Discussion Up: Controlled Experiments Previous: Varying the Predictor
Peter Stone 2003-09-24