Studying Complex Game Systems Part 4: Advanced Artificial Intelligence in Games

Our next topic for the Complex Game Systems subject was advanced artificial intelligence. In this topic we scratched the surface of advanced methods of developing artificial intelligence in games and other industries, including neural networks and genetic algorithms.

Neural networking is a technique applied to machine learning that aims to mimic the human brain. Neural networks are made up of input, hidden and output nodes or ‘neurons’, connected to each other through synapses in that particular order. Input neurons receive information and input from the surrounding environment, triggering it. When an input neuron is triggered, synapses apply a weighted modifier to an Action Potential (AP) and sends it onto the hidden neurons. If the total value of the AP a hidden neuron receives is greater than its threshold, the hidden neurons passes on its own AP to other hidden or output neurons. An output neuron is triggered when the amount of AP it receives is greater than its threshold. The information gathered and the conclusion drawn from it are then outputted. The process of implementing a neural network for AI purposes involves its training. These networks are trained to complete certain tasks over time, learning with each new piece of data or adjustment. In addition to game development, neural networks are used in many areas, including marketing, medicine, sales and finance. The different forms of neural networks have also been used for things like facial recognition, self-driving cars and sales forecasting.

Another machine learning technique used is the genetic algorithm. Genetic algorithms fall under the category of evolutionary algorithms and aim to mimic the process of natural selection. This is done by consistently attempting and re-evaluating solutions to a problem or task. The most viable solutions are combined to make new ones and the unsuccessful solutions are abandoned. In the same vein, the population of agents are also evaluated based on their fitness to survive. Each agent’s fitness score effects their chances of being selected to pass on their genome to the next generation. This is done via a process called Crossover, where the genome of two selected agents is combined to create a new sibling. A regulated mutation rate is then applied to the new sibling’s genome, to insure the possibility of new behaviour and approaches to the populations previous solutions. This new generation then goes on to replace the original population. Other than game development, genetic algorithms are also used for screening in fields such as radiology, cardiology and oncology.

References:

Neural Networks Lecture

Genetic Algorithms Lecture

Ghaheri, A., Shoar, S., Naderan, M. and Hoseini, S., 2015. He Applications Of Genetic Algorithms In Medicine. [online] www.ncbi.nlm.nih.gov. Available at: <https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4678452/> [Accessed 7 June 2020].

Shah, J., 2017. Neural Networks For Beginners: Popular Types And Applications. [online]

Medium. Available at: <https://blog.statsbot.co/neural-networks-for-beginners-d99f2235efca?gi=7e53721a1bef#:~:text=Today%2C%20neural%20networks%20are%20used,data%2C%20and%20natural%20language%20understanding.> [Accessed 7 June 2020].