Grand Text Auto

Building expert-level artificial intelligence for real-time strategy games remains an open research challenge. StarCraft in particular provides an excellent environment for AI research, because the game has many real-world properties and is played at an extremely competitive level. It is also an environment in which human decision making can be observed, emulated, and evaluated. During gameplay, professional players demonstrate a broad range of reasoning capabilities including estimation, anticipation, and adaptation.

While I have been advocating the use of reactive planning for over a year now, there is often a large amount of middleware between a game environment and the reactive planning agent that needs to be defined in order to make use of ABL in games. The goal of this article is to provide a tutorial for interfacing a game environment with a simple ABL agent.

Madden NFL 11 – EA Sports

Game telemetry is being used both during development and post release. One of the most exciting applications of this work is the use of game telemetry to support the game design process. Game telemetry analysis can help a designer answer the following questions:

• How do players interact with the game?
• Which features, modes, and content are players experiencing?
• Why do players quit playing the game?

This year the AAAI Fall Symposium Series included a track organized by Pat Langley on Advances in Cognitive Systems. Pat identified the following objective for the symposium:

Pursue the initial goals of artificial intelligence and cognitive science: To explain intelligence in computational terms and reproduce the entire range of human cognitive abilities in computational artifacts.

Pat’s motivation for re-evaluating the goals of artificial intelligence is presented in a recent editorial in Machine Learning, while my motivation for attending the symposium was to discuss how to build integrated, heterogeneous agents. The introduction to the symposium was a discussion of why AI research has gone astray. Pat identified the following issues in his talk:

Dennis "Thresh" Fong

Dennis “Thresh” Fong, a retired pro gamer and founder of Raptr and XFire, participated in an exhibition match against EISBot during a visit to UC Santa Cruz today. While Dennis had not played StarCraft: Brood War for several years, he provided EISBot with an excellent challenge. Dennis played as Zerg (orange) and the outcome of the match is shown in the video below:

The StarCraft AI competition introduced at AIIDE 2010 will be part of the AIIDE 2011 program. Last year the main event was won by UC Berkeley’s team, which showed how a computer opponent could be used to destroy enemies. This year, we expect to see even more sophisticated agents. Competition details are available here.

Alex and I will be discussing the current state of RTS research and the StarCraft AI Competition today at 13:00 PST. https://my.dimdim.com/aigamedev/

The AIIDE 2010 StarCraft AI Competition has come to a close. The challenge given to competitors was to build the best performing bot for an immensely popular, commercial game. The competition consisted of four tournaments of varying complexity. This was the first year the competition was held and it turned out to be a success. Even though no prizes were offered, twenty-eight teams participated in the competition. My presentation on the competition provides an overview of the  participants and results.

Recently, there has been increased interest in building games that dynamically adapt to players. One of the common approaches to building adaptive games is dynamic difficulty adjustment. However, most of these approaches are limited to parameter tweaking such as adjusting weapon strength or reducing spawning times, and do not modify levels in response to difficulty adjustment. My system attempts to overcome this limitation by incorporating parameter tweaking into procedural content generation. The system creates new levels on the fly in response to the current performance of the player.

I recently participated in the CIG 2010 Mario level generation competition. My submission utilizes a multi-pass approach to level generation in which the system iterates through the level several times, placing different types of objects during each pass. During each pass through the level, a subset of each object type has a specific probability of being added to the level. The result is a computationally efficient approach to generating a large space of randomized levels.

A level created by the probabilistic multi-pass generator

The generation process consists of several phases, which place additional object types in the level. The following passes occur during generation:

Submission for the StarCraft AI Competition is now open. Complete details are provided at the submission site.

New Scientist is running an article about the use of data mining in computer games. The article focuses on research being presented at the IEEE Conference on Computational Intelligence and Games (CIG 2010). Catch live coverage of the conference here.

Command your space bovine!

In space, social interactions are sparse. Space Cow Clicker overcomes this problem. In this parody of a satire, you command a battlecruiser (Space Cow) in an epic battle  to click enemy units. The first player to click all of the opponent forces wins: To click is natural, to command is Bovine!

In April, I blogged about adding chat capabilities to EISBot, with the goal of achieving the Eliza effect in StarCraft.  Nicholas Carr responded to my post, criticizing my approach:

The sure way to distinguish the computer’s messages from the human’s is to recognize that the computer has a rather sentimental attachment to the apostrophe and the comma.

My dissertation work aims to build intelligent game AI by learning from replays. However, the major limitation of this approach is that the AI cannot be developed until a large number of players have first played the game. Fortunately, large-scale beta testing is becoming more popular for games. These beta cycles result in mountains of data that can be used to build AI for games. One of the most notorious beta releases is Blizzard Entertainment’s StarCraft 2, which ran from February 17 to June 7. The only AI provided with this release was “very easy”, which means that Blizzard may be analyzing how players actually play the game before finalizing the AI.

‘Kasumi’s Stolen Memory’ is a DLC mission for Mass Effect 2 that adds a new perspective to gameplay in the Mass Effect series. While the DLC contains the formulaic loyalty mission for the new character, it also puts Commander Shepard in a new role in which the player interacts in a formal social setting. Shepard’s mission is to assist Kasumi in infiltrating an extravagant party in order to reclaim Kasumi’s personal artifact contained in the vault of the party’s host. Part of the DLC is a new formal wardrobe for Shepard (pictured below), that while only providing a reskinning, changed my perspective of the character. Playing through this mission reminded me of the scene from the interactive drama Heavy Rain in which the journalist (Madison Paige) needs to infiltrate a nightclub to acquire information from the owner. After drawing this comparison, I found myself asking the question: Can Mass Effect 2 be considered an interactive drama? Can the player have meaningful participation in the development of the plot in an action game?

I recently ported Infinite Mario to the Android platform in order to evaluate the use of procedural content generation in mobile devices. My goal was to develop a casual platform game, where players can jump right into the action without worrying about making it to the next save point. Procedural content generation is used to provide the player with short, unique gameplay sessions. The player’s goal in Mario Forever is simply to complete a level. The major challenge in this game is learning to defeat levels you have never encountered before.

Mario Forever

Once again, the IEEE conference on Computational Intelligence and Games (CIG 2010) is hosting a Mario competition. This time, they’ve added a level generation track to the competition. The goal is to procedurally generate Mario levels that are entertaining. EIS is working on an entry and here’s what we have so far:

I’ve made a few modifications to the game engine to explore a new direction in platformers. Each time the player loses, a new level is generated on the fly. The idea is to prevent the player from getting frustrated by always  presenting the player with a new level. Rather than frustrate the player, the system provides the player with unlimited fun levels. You can try out our level generator by downloading it here and running the Jar.

Games are increasingly making use of gestures as a way for players to interact in game worlds. While enabling players to perform a large number of actions in the game world, the use of gestures in games can also confuse players accustomed to conventional controls. Bob Mitchell’s recent talk at UC Santa Cruz discussed developing games for 2020 and one of the points he made was that gestures are going to become prevalent in games due to the introduction of new interfaces. However, it is first necessary to build standard definitions of gestures in order for players to build expectations of how to interact with games.

One of my research goals is to build an agent that mimics human gameplay. To achieve this challenging goal, I have implemented chat functionality in EISBot. Once every thirty seconds, EISBot randomly selects a message from a pool of 75 messages and sends it to the console. The results can be quite convincing:

Here is the complete transcript of messages:

Play against the EISBot! Download here: http://eis.ucsc.edu/sites/default/files/eisbot.zip

THE EISBot in a Protoss versus Terran match

So far, I’ve tested the bot versus the avid StarCraft fanbase. It pulls out a 10% win rate versus hardcore gamers.  Can you beat it? Play it and send me the replay!!!

Note: playing against the EISBot requires two machines capable of playing StarCraft.

I developed a version of EISBot that plays a specific strategy, known as a 10-15 gate rush. The build was recently made famous by Nony. It is a Protoss build with the goal of harassing your opponent with ranged units as fast as possible and is most commonly used against Terran opponents. This strategy requires a large number of micro-management actions to maximize the efficiency of each unit. This is achieved by attacking enemy units and then backing off.

I tested the bot against several human players on the International Cyber Cup and was surprised to discover that the bot achieved a win rate of over 30% when tested in 40 matches.