Evaluating and Enhancing Gameplay Behavioural Expressivity of Planning-Playing Artificial Intelligence for Automatic Playtesting
Abstract
This thesis focuses on game-playing Artificial Intelligence (AI) for advancing fully-automated play-testing using Statistical Forward Planning (SFP) algo- rithms like Monte Carlo Tree Search (MCTS) and Rolling Horizon Evolutionary Algorithm (RHEA). First we provide a general methodology to analyse and compare the SFP’s decision-making in General Video Game Playing. This analysis highlighted that algorithms are mostly influenced by the reward signal and less by their parameters. For improving fully-automated play-testing, they need more subtle and tunable reward functions. The Game AI framework Rinascimento was developed, based on the board- game Splendor, to provide richer information to the players and move beyond the concepts of state-value and state-action-value functions. Rinascimento is modular, highly-efficient, forward-model-equipped and statistics-oriented and it can run any Splendor-like game, the game-parameters can be freely changed. Several old and novel SFP algorithms were implemented and we ran experiments to find their best configuration. A novel type of reward, Event value Functions (EF) were developed and shown to provide stronger and more controllable agents than conventional state-value functions. To explore behavioural expressivity of AI players we ran Multi-dimensional Archive of Phenotypic Elites (MAP-Elites) experiments. Searching the AI’s hy- perparameter space (algorithm and reward function parameters) they illuminate the behavioural space of several game configurations. EFs induce remarkable breadth of behaviours and skill depth compared to state-value functions. A pub- lic online tool helps visualising the multi-dimensional spaces and can compare two experiments highlighting coverage and performance differences. Finally, we designed several experiments for the fully-automated search of game variants using either Relative Algorithm Performance Profiling (RAPP) or Behavioural Evaluation (BEv) fitnesses. BEv can find games targeting a specific behavioural space making it more expensive but higher-quality output than RAPP. This thesis has provided a novel, efficient, reliable and general pipeline for comparing AI-players performance that, thanks to more expressive agents, brought richer, faster and more thorough automated play-testing deployable for automatic search of game variants.
Authors
Bravi, ICollections
- Theses [4190]