Generating Levels Using Player Experience

Extracting the race track's spline from the Voronoi diagram starts by randomly selecting an edge. Then creating a list of edges (referred to as neighbouring edges) which are in direct connection to our selected edge. A neighbour edge is randomly selected from this list and the vertex point between our edge and neighbour edge is stored as a point within the race track's spline. That neighbour edge is now the selected edge, repeating this process until we have returned to our initial edge.

This process ensures that the race track is a complete circuit with no segments overlapping as an explored edge can not be reused. On top of this by applying Perline Noise to the spline it was possible to add variation to the race track's Y-elevation throughout the spline.  

A sped-up simulation of a race using the race track is performed. A race involves having four cars on the track known as agents, with one of these agents being the player agent. The player agent will be logging its data (speed, visible opponent’s speed, standing, etc) and passing this data through the human experience intensity model to obtain a player experience (PX) value between 0 and 1. After this conversion the PX value is placed in a temporary array and the lists of data are wiped.

The player agent is logging data every FixedUpdate (0.02 seconds). To avoid memory overflow a PX value is obtained when the data lists reach 50 elements, making it possible to frequently clear the data lists. At the end of each segment (gene) a final PX conversion if performed to use up the remaining logged data, followed by calculating the average PX for that segment based on the list of PX values obtained during that segment. The human experience intensity model functions by comparing the input data to the list of human play traces and outputting the experience intensity value of the human play trace which most closely resembles the current game state.

There are three scenarios in which a simulation ends. Either the player agent has completed one lap, the player agent got stuck or time ran out. An agent is considered stuck when it has not moved much distance within several frames. In this scenario, the agent is teleported to a previous waypoint. After doing this several times it is concluded that there must be an unplayable segment. The time limit is simply how long the race track is allowed to be active before being forced to end. The time limit takes into account the length of the race track and is a generous amount of time. This is more prevent any physics-based bugs such as the agent being out of bounds from halting the GA.

Once the race is concluded the program moves on to calculate the fitness score. In the scenario that a segment was not assigned an average PX value, it will be assigned a value of -1, which is unobtained. The fitness score starts at 0 and is subtracted by the difference between the simulated average PX and the target PX for that segment. 

|Target PX − Simulated PX| = Amount To Subtract From Fitness 

There are three possible scenarios for a race to end. Either all the player agents complete one lap, all the player agents get stuck or time runs out. An agent is considered stuck when it has moved much within several frames. In this scenario, the agent is teleported to a previous waypoint. After doing this several times in the same race it is concluded that there must be an unplayable segment. The time limit is simply how long the race track is allowed to be active before being forced to end. The time limit takes into account the length of the race track and is a generous amount of time which is intended to catch any physics-based bugs which may result in the agent being out of bounds and halting the GA from progressing. 

With the race being concluded it is now possible to obtain the fitness of the race track using the array of average PX. In the scenario that the array of average PX is not equal to the number of segments in the track, it is concluded that the agent failed a segment and the ones that followed, therefore these segments are assigned an unobtained average PX of -1. The fitness starts at 0 and is subtracted by the difference between the simulated average PX and the target PX for that segment. 

|Target PX − Simulated PX| = Amount To Subtract From Fitness 

The Genetic Operators of selection, crossover and mutation are applied to the current population. This results in a new population of race tracks based on the data from the previous generation. This seeks to optimise the race tracks towards providing the best output, along with exploring search space.