Research: A FPS which recognizes your actions to adapt itself
You buy a new game and the first step you meet with your game play is the “difficulty” setting. You pick the best estimated difficulty and start playing, to only realize later that this setting is either too easy or too tough. Ever felt, “I wish there was a setting between easy and medium”? Can walkthroughs be made obsolete?
Few games (Splinter Cell: Pandora Tomorrow, Lego Star Wars II) have tried to handle situations like above by making the game harder according to the gamers success, however they do not respond to specific advanced skill sets possesed by the player or the lack of.
Most non player characters (NPCs) are built with a very limited set of capabilities and navigation. A different or unconventional approach used by a gamer often leaves NPCs lost. For example, in several FPS, advanced gamers would seek a good & safe vantage point which attracts a good number of opponents (NPCs) to take them down off-guard (spawn camping, sniping etc). Gamers engage in this as they see a lot of opponents in the vicinity, they have got a good vantage point and they are sure that the AI of the game would keep sending almost all of the NPCs in the vicinity (in most cases through the same path). How interesting would the game be, if the AI was able to detect that the gamer now has a vantage point, lure him into a different/difficult location rather than sending more NPCs. Considering a novice gamer who lands up with no ammunition very quickly, if the AI could see that the gamer was wasting way too much ammunition, control the power of the autonomous weapons to limit the ammunition wastage when there is no NPC in clear view and thus lessen the frustration the gamer would experience.
Researchers from UK have now tested a prototype FPS which can adapt certain aspects of the game to best suit the player to endure the best possible engagement. To examine such a system, the researchers built a test game (with no adaptive game play) using the “Unreal Editor”. The game was then tested for correct functionality by 30 students of mixed sex and gaming abilities. A different random pool of 16 gamers with varying gaming abilities were then used to conduct the actual experiment. The test game included a total of 18 rooms, with several possible routes to reach the end point, safe passages, secret short routes, elevators, stairs etc. First two rooms of the game were used for making the players comfortable with the game (invincible mode). The gamers were then allowed to try completing the game, with a maximum of three attempts. Event handlers were implemented to write strings to text files when gamers visited certain areas in the map (like a safe passage or a secret path), interacted with game elements such as doors, weapons etc, killed a player, were being injured etc.
After all the 16 gamers completed, the data collected was analyzed. Only five players successfully completed the game within three attempts. Data showed that there were eight expert gamers, six intermediate games and two amateur gamers. Researchers then made several observations about the top five and least five players. Briefly, the top five completed the game, visited most of the rooms (mean=13.4), a few used secret and/or short routes, none of them used the most linear route, four of them used only one life, a few picked up extra weapons and/or health. Of the least five, all of them failed to complete the game (3 kills, 2 suicides), never used any secret/short paths or weapon/health pickups, all of them used either a linear or a linear with slightly modified path and only visited an average of 4.2 rooms.
Novice players exhibited a few typical characteristics such as: i. firing at a close vicinity to a stationary object like a wall or door (sometimes even closing doors) and thus resulting in death (suicide) or massive health loss, ii. spending lots of ammunition to cause very little to no effect on the NPCs and thus landing in a situation with no ammunition to further combat, and iii. a significant drop in gamer health level in a short span of time, due to being not able to avoid damage caused by the NPCs.
On the converse, experienced players exhibited the following characteristics: i. defeating multiple NPCs within a very short span of time, ii. usage of killzones, where they simply stand and snipe from a specific location, iii. adopting tactics like passing through corridors to block NPC fire and to launch counterattacks from, and iv. choosing rooms for launching attacks which tactically give them an advantage and only navigate into rooms which are absolutely essential to complete the game.
Using the above observations and Finite State Machine based approach, different adaptive strategies were developed. Some of the strategies for the advanced players were: i. when the gamer uses a particular safe passage/corridor, the geometry of the path further is dynamically modified to make it tough for the gamer to avoid the NPCs and also a few NPCs are positioned further down the path. ii. a kill zone counter mechanism where the system measures the number of kills within a short time period with the gamer maneuvering very little, once detected NPCs are dispatched through an alternative route. iii. a lure mechanism to go along with the kill zone counter, where the algorithm lures the gamers from his high safety point to a point where NPCs have an advantage to create an ambush like situation. Strategies for the amateur players were: i. automatic blast radius adjustment mechanism gets activated when gamer repeatedly instigates damage to himself due to fire from his weapons. ii. automatic weapon directional control which minimizes the wastage of ammunition when detected iii. automatic NPC skill reduction mechanism, activated when there is a sudden heavy loss of health due to the gamers inability to handle NPCs attack. The FSM also measured the overall skill level of the gamers.
Upon implementing the above adaptations, 36 new participants of mixed sex and gaming abilities were recruited and were asked to play the game, with a maximum of three attempts. Statistical analysis revealed that the adaptations impacted both amateur and expert players significantly. Some of the most important observations were: 61% of gamers were able to finish the game as compared to 31% before the adaptations, 67% of the gamers used 12 or more rooms as compared to only 25% before adaptations, most (36% of gamers) of the deaths were in room 17 (close to finish) rather than in room 4 (25%) and 6 (36%) when not adapted, amateur player deaths decreased from 56% to 28%. It is now highly possible that one assumes that the “game got easy”. Further analysis on only the expert gamers showed that: 39% of expert gamers were killed in the adaptive game as compared to 6% before.
None of the expert players ever activated the mechanisms built for amateur players while only one amateur player activated one advanced mechanism (“modify geometry”) due to his exhibit of advanced skills at that point of the game. The researchers identify that, there is lots of room for further work like: interconnecting the various mechanisms and evaluating the gamers on a continuous basis, so as to activate certain triggers even before the gamer enters a region (example: if 3 players trigger a exact same set of mechanisms and completed a game, then the AI should recognize the statistical likelihood and further adapt the game), amateur players could be introduced to adaptive audio warnings, for expert gamers NPCs could be made to roam anywhere within a level rather than just one or few rooms – using a navigational game tree of the game space.
Although this has been tested with FPS, the same approach can be used for other genres such as role playing, vehicle simulations etc.
Journal Reference: International Journal of Computer Games Technology
