Researchers at Essex University hope to give video game characters the power to think for themselves instead of relying on software instructions.
A new research project will attempt to make computer games more realistic, using a recently developed type of computer algorithm. The programming was first used to simulate board games, but is also being applied to complex applications such as logistics and energy management.
The algorithm simulates the effects of a large number of possible actions in order for the computer to decide what to do. This could allow computer-controlled characters to do things that they have not been specifically programmed to do.
‘You give it the model and it essentially works out what to do,’ Essex professor of computer science Simon Lucas told The Engineer.
‘You could be playing a video game and have the impression you’re up against some real intelligence that can react to novel situations that have never really been seen before.’
The research will use a version of Monte Carlo Tree Search (MCTS), a class of algorithms developed in 2006 and most commonly used to play the strategy board game Go.
MCTS works by repeatedly and randomly simulating what will happen if it makes a series of moves in the game, and uses this information to calculate the value of individual moves and decide which one to make. For Go, MCTS could typically simulate around 100,000 games or more before making each move.
This differs from traditional forms of computer game artificial intelligence that examine every possible move but don’t look as far ahead in the game. They also need an evaluation program with expert knowledge of the game to decide what is a good move.
In more complex situations with more variables, these evaluation programs don’t work very well or are too hard to build. So MCTS enables the computer to look far enough ahead to find the ultimate likely outcome of a move.
‘You’re trading the breadth for depth,’ said Lucas. ‘It has a very nice way of balancing this trade-off between moves that appear good and ones that, at first glance, didn’t appear to be good but the estimate wasn’t very reliable.’
The appeal for game designers is that MCTS could be dropped onto any system and generate intelligent, creative actions, said Jeff Rollason, chief executive officer of game company − and one of the project’s commercial partners − AI Factory.
‘As a gamer, I’d much prefer if the non-player character really did seem to be getting the better of me − and not just through physical skills in the game but by making better decisions,’ he said.
As part of the project, funded by the EPSRC, collaborators at Bradford University and Imperial College London are applying the algorithms to computer card games and computer-generated games, respectively.
MCTS also has potential uses in complex real-world applications such as energy management, where the algorithms could help make the best use of intermittent renewable power while meeting changing levels of demand.
‘MCTS decides which power plants should be switched on or off, which quantity of water should be used for producing electricity today and so on,’ said Olivier Teytaud of The French National Institute for Research in Computer Science and Control (INRIA), which is studying the issue.
‘This studies are in simulation only for the moment…but we have serious hope that the MCTS becomes the standard tool in the future for these problems,’ he said.