Maciej Świechowski, Ph.D.

The main result of research carried out within the MPD framework was related to design and implementation of a complete General Game Playing (GGP) system. This system, also referred to as the GGP player, is capable of playing any finite, deterministic and perfect-information game written in a Game Description Language (GDL).

The resulting system incorporates a few novel methods in the area of multi-game playing AI systems:

  1. Application of knowledge in Monte Carlo Tree Search (MCTS) used together with the Upper Confidence Applied to Trees (UCT) algorithm by means of strategies (heuristic simulation policies) which are assigned to players in a simulation phase of the MCTS algorithm to bias the, otherwise random, search. In this approach, the knowledge is used in a light-way only to optimize the search which is still statistically-based. 
  2. Introduction of three novel heuristics in the GGP framework. Two of them are based on knowledge extraction from logical rules. Some known heuristics were also revisited and incorporated into the system.
  3. Self-adaptation mechanism for the strategies to select the right strategy for a game based on its dynamically evaluated statistical performance. 
  4. Design and implementation of a fast GDL interpreter. The interpreter is used to speed up Monte Carlo simulations and enables logical reasoning based on GDL rules.
  5. Proposal of a novel method of parallel execution of the player’s code to harness power of multi-core environments.

Moreover, the player called MINI-Player took part in 2012, 2013 and 2014 International GGP Championships and competed with the top players from all over the world. In 2012 and 2014, it ended up in 7th place out of 11 and 49 players respectively, which is a reasonably good result.

The main algorithm applied in the GGP system was successfully adopted in another problem strengthening premises about its universality. The problem was Dynamic Capacitated Vehicle Routing Problem (DVRP) with dynamic traffic jams which may be regarded as a combination of a generalized Travelling Salesman Problem, a generalized Discrete Knapsack Problem and dynamic path optimization.

Mr. Maciej Świechowski published the results of his work in many papers (cf., here).