: research

Play balancing massively multi-player games

Summer 2000
With Jessica Hodgins

Computer Animation Lab, Graphics and Visualization Unit
Georgia Institute of Technology

Summary

Over the past few years a new genre of video games has emerged: massively multi-player games. These new games, Ultima On-line, Everquest, and Asheron's Call, have the potential to add a significant new demographic to the computer game industry because the user-population is older and more gender-balanced than the young, male population that has made twitch games the dominant genre. These games also force a new economic model on the industry because the user must play $10/month to attach to the server as well as paying the up-front $50 cost of the game. The popularity of these games is significant and growing. These games all operate with multiple servers, each of which handles thousands of users during peak times. These games have also created a secondary economy; powerful characters and items of clothing are selling on Ebay for prices ranging from $50-$500 with more than twenty different auctions open at any given time.

Debugging and play balancing these games has proven to be quite difficult. The users of the early games in this new genre behaved in ways that the designers and testers did not predict in advance. For example, the users were much more willing to engage in player kill and this behavior has had a significant effect on the atmosphere of the game and discouraged many new users. The users have also discovered sequences of actions that form "money plays." These plays are sometimes so lucrative that they ruin the game play and have to be made illegal. For example, players will hang around the spot where a particularly valuable pair of boots are known to spawn hoping to pick up a pair rather exploring the world as the designer intended. In current implementations, these problems are dealt with in a heavy-handed fashion: servers have been set up where player kill is not allowed and the valuable boots now randomly spawn in a set of locations and may be eliminated altogether. Play balancing these games more completely in advance of their release would be a far preferable solution because intervention after the fact destroys the players' sense of presence in the virtual world.

During the summer, we plan to explore whether optimization techniques in general and design galleries [1] in particular can be used to play balance games. We will start by exploring permutations to more traditional games such as chess, Monopoly, and Risk. These games have withstood the test of time and therefore are presumably well balanced. We will change the rules for a game (kings can now move like knights in chess, for example) and test the new game for playability. How long does the game last? Does one side have a significant advantage over the other? Does one style of play tend to win? We plan to generate AI algorithms with particular strategies to play test the games. Either a hand-tuned evaluation function or human inspection could be used to determine which of the permutations appear to create a fun game. In the longer term, we plan to experiment with play balancing massively multi-player games such as Asheron's Call. The developers of these games have already built fast-play modes in which the games are driven by scripts without graphics to facilitate debugging. The games are also instrumented to create log files of actions leading to the possibility of building data-driven models of a user's behaviors to seed the performance of the AI algorithms.

References

[1] J. Marks, B. Andalman, P. A. Beardsley, W. Freeman, S. Gibson, J. Hodgins, T. Kang, B. Mirtich, H. Pfister, W. Ruml, K. Ryall, J. Seims, and S. Shieber. Design galleries: A general approach to setting parameters for computer graphics and animation. Proceedings of SIGGRAPH 97, pages 389-400, August 1997. ISBN 0-89791-896-7. Held in Los Angeles, California.

Phase I: Monopoly

Parker Brothers introduced Monopoly in 1936, and the game is still wildly popular today. Its longevity suggests a well-balanced design, and the game is thus a good choice for analysis. Though the game is highly dependent on luck, every player has pet strategies. Whether any strategies statistically increase a player's success is the subject of the first phase of this project. Additionally, we will evaluate the effect of changing typical game parameters, such as the amount of cash players get at the beginning of the game, and how much they get for passing Go.

Game Engine
Using a Monopoly game engine written by student Cem Cebenoyan that simulates computer-driven players and prints output to a log file, we are analyzing the effects of different player "personalities" on game outcome. Input to the engine includes: number of iterations, number of players, typical game data such as property rents, and a list of AI strategies for each player.

Unlike the standard board game, the computer model has the following modifications:

* An artificial ceiling of $100,000. When any player hits this limit, the game is declared winnerless and terminated. The ceiling is necessary to prevent the infinite games created by particular strategies.

* No building shortage. Players still build up to one hotel per property, but the bank never runs out of materials.

* Simplified auctions. To avoid bankruptcy, players sell their properties for full face value to the debt collector (either another player or the bank).

These modifications lead to simpler bookkeeping, though they reduce the realism of the simulation.

Player Strategies
Each player has three main factors to his or her personality: property buying tactics, trading preferences, and house building strategies. Correspondingly, each AI personality has three parts, which are subdivided into base preferences and modifiers. For each of the three parts, a player has one base preference and a binary value (y/n) for each of the modifiers.

Property Buying
Base Preferences:

Modifiers:

Property Trading
Base Preferences:

Modifiers:

House Building
Base Preferences:

Modifier:

Experiment
To test these different personalities, we ran a series of four-player trials, with one trial consisting of 10,000 games. By running a large number of iterations, the role of luck was minimized, accounting for approximately 1% error. Matching four identical players, knowing that each should win 25% of the time, we saw the players win between 24 and 26% of the games.

Rather than exhaustively testing all combinations of strategies against each other, we used a form of interactive evaluation, using the results of previous trials to chose future matches. Each trial generally focused on the effect of a single trait: a base preference or a modifier. The trials consisted of three types of matches:


Equal Pairs
Used to compare the effect of a particular trait on otherwise identical players. Players 0, 1, 2, and 3 have exactly the same preferences, but players 0 and 1 have the desired trait turned on, while players 2 and 3 have it turned off. The players' performance is judged against a 25% level, the average player's success.

Unequal Pairs
Used to view the effect of a particular trait across several player profiles. Players 0 and 1 have identical profiles, except that player 0 has the desired trait turned on, while player 1 has it turned off. Players 2 and 3 have identical profiles, though they might be completely different from the profiles of players 0 and 1, and again, player 2 has the trait turned on and player 3 has it turned off. Many sets of unequal pairs are played to reveal the change in magnitude of a trait's influence.


Quartets
Used to analyze the relative effects of non-binary traits, such as base preferences. The participants have identical profiles except each has a different base preference (i.e. never trade, always trade, trade if wealthier than opponent, random trade). The results indicate which players should be paired for future matches.

Results

Property Buying

Always buy vs. Leave padding (but also buy matches and pieces opponents want)

Players that always bought properties performed slightly (3%) worse than players that left $200 padding in their accounts, dipping below that level only for matching pieces and pieces opponents needed to make complete sets.


Effect of changing padding

Key:
Player 0: Always buy
Player 1: Leave padding, but buy matches
Player 2: Leave padding, but buy matches and pieces opponents want
Player 3: Random buy

Changing the padding value dramatically altered the success of particular strategies. At a relatively benign $200, all players fared about the same, though the player that always bought properties performed slightly below the others. With the padding set to $1000, the players that aimed to keep this amount in their accounts performed much better than the player that always bought and the random buyer. At $1250, the results leveled off again, but at a prohibitively high $1500 (the amount the players received at the start of the game), the players that kept padding in their accounts began to lose much more often.

Other property buying results

Property Trading

Never trading

The most dramatic effect of any strategy was the detrimental effect of never trading with opponents. Players that rigidly refused to trade with others won only a small (4%) portion of the games. Additionally, many (6%) of the games would have run infinitely.

Most likely the infinite games were caused by the lack of complete sets. Since players must own an entire set for development, rents remained negligible and players continued to lap Go.

Not trading pieces opponents want

Players that refused to make trades that helped opponents make complete sets also fared below average. Players that always traded performed 8% better than those that held back properties that opponents wanted.

Perhaps a more complicated, flexible AI strategy could be devised to retain pieces opponents wanted and still outperform strategies that always trade pieces.

Other property trading results

House Building

Build up one set at a time vs. Spread across board

Players that completely developed one set before building houses on another performed 2% better than players that spread houses across all sets.

Building houses where other players were likely to land gave no appreciable benefit.

Composite "best" strategy

The following traits increased players' chances of winning slightly. However, the benefit is too small to notice in only a few games. So, for a human player, this strategy is not appreciably better.

Conclusions

Though there are a few detrimental tactics (i.e., refusing to trade), there are no exceptionally beneficial strategies for a human playing a limited number of games. Thus, Monopoly is a highly balanced game.

Future Work
To further explore the game space, we will investigate other avenues, including:

Finally, we will move the local problem back to the wider scope of games in general. We will investigate other, less luck-bound games such as Risk and chess. Finally, we will apply these principles to the new massively multi-player server games, hoping to remedy game imbalances with successful techniques from traditional games.

Also see
Other research projects