Reinforcement Learning of Intelligent Characters in Fighting Action Games

Advanced Dynamic Scripting for Fighting Game AI

“…A classic approach, which we found in many articles, was the use of a combination of reinforcement learning and neural networks [1,5,3]. The techniques, applied in these articles, work well for very simple games but need long training phases and are, therefore, not appropriate for real-time games.…”

Comparison of AI Techniques for Fighting Action Games - Genetic Algorithms/Neural Networks/Evolutionary Neural Networks

“…One characteristic of board games is that you must be aware of the overall situation of the pieces on the board and determine their movement. Recently, there have been studies to apply neural networks to fighting action games [7,8,9]. These studies used the action and step of the opponent character and the distance between characters as the input for neural networks, and the difference of scores resulting from the actions of two characters as the reinforcement value so as to make the intelligent characters learn whether or not their current action is appropriate.…”

“…These studies used the action and step of the opponent character and the distance between characters as the input for neural networks, and the difference of scores resulting from the actions of two characters as the reinforcement value so as to make the intelligent characters learn whether or not their current action is appropriate. [7,8,9] expressed the neural network to represent intelligent characters as Figure 1. In Figure 1, input is the information related to the opponent character, and PA(t) indicates the action of the opponent character at time "t", while T indicates the progress level of a particular action, and D indicates the distance between the intelligent character and the opponent character.…”

“…In such genres as fighting action games, it is more important to make the characters intelligent so that they can cope appropriately according to the location and action of surrounding opponents, rather than to tune up the overall strategy of the game. Due to this reason, many have carried out studies to implement intelligent characters in fighting action games [7,8,9,10]. They used genetic algorithms, neural networks, and evolutionary neural networks to create intelligent characters for fighting action games.…”

“…Studies using genetic algorithms evolved intelligent characters by including the distance between characters, the action and step of the opponent character, and past actions in the chromosomes that express intelligent characters [10]. Studies using neural networks used the action and step of the opponent character and the distance between characters as the input for neural networks, and the difference of scores resulting from the actions of two characters as the reinforcement value so as to make the intelligent characters learn whether or not their current action is appropriate [8,9]. However, because neural networks that use the error backpropagation algorithm have such shortcomings as falling in local minima and slow convergence speed, methods to apply evolutionary neural networks were studied [7].…”