1. Nicholas Mifsud

2.  Behaviour Trees (BT) proposed as an improvement over Finite State Machines (FSM)  BTs are simple to design, implement, easily scalable, modular and reusable  BTs provide a hierarchical way of organising behaviours in a descending order of complexity (broad tasks on top, sub-tasks at the bottom)  Each tree has one defined high level behaviour, multiple trees are used to define an AI-bot

3.  Nodes are divided in two categories – ◦ Control Nodes  Drive execution flow through the tree, deciding which nodes to execute next  Sequence – execute children left to right until one fails (logical AND)  Selector – execute children until one fails (logical OR) ◦ Leaf Nodes  Conditions – query the game state (left most)  Actions – carry out specific tasks (sequence of actions follow conditions)

4.  Lim et al. investigate feasibility of applying evolutionary techniques to develop BTs for competitive AI-bots  Authors give a brief outline of ◦ DEFCON ◦ Relevance of BTs ◦ Four different fitness functions ◦ Experiments and Results ◦ Future Work

5.  Multiplayer Real-Time Strategy game  Six world territories  Players have set of units and structures at their disposal  Default AI-bot uses a deterministic FSM consisting of 5 sets and follows them in sequence  AI-bot was given all functionalitites to play the game (makes use of DEFCON API)  Decisions performed randomly

6.  Original set of trees were hand defined in order to cater for all possible actions  Handcrafted trees then used as a basis to produce more complex behaviours  Develop AI-bot behaviour by following an evolution algorithm (genetic operators + fitness functions)  Evolved behaviour trees for individual behaviours and combined the best performing trees

7.  Tree structures naturally allow genetic operators to modify the behaviours  Crossover on branches  Mutation on nodes ◦ Random mutation ◦ Incremental mutation  Add a new branch to a behaviour ◦ Point mutation  Change the spawn point of a structure/unit

8.  Defence potential – total number of air units destroyed in a given game  Uncovered enemy units – number of enemy sea units uncovered  Fleet Actions – number of enemy buildings uncovered, buildings destroyed and number of units destroyed  Timing of attacks – difference between final end scores as an overall tally (large difference would mean convincing win)

9.  Four different experiments run each to develop a different fitness described above  Each population contained 100 individuals  Each experiment evolved between 80 to 250 generations  Mutation rate set to 5%

10.  All results indicate the mean of each fitness function increased as more generations were produced  Higher percentage win rate

11.  AI-bot constructed with a controller that used the best trees evolved for each of the four behaviours  Played 200 games against the default AI-bot  Won 55% of the time  Games that it lost, was by a very low margin  Before evolving the behaviour trees it only won 3%

12.  By using evolutionary computing and BTs, able to defeat a handcrafted AI-bot  Mean fitness reached a plateau indicating more generations may be redundant  Raises question whether other techniques could be coupled with this process  Not all possibilities of game explored (only two territories and only 4 tasks were considered)

13.  Perez et al. also investigate the applicability of Genetic Programming to evolve BTs  Authors make reference to DEFCON paper discussed above and state that their work may be extended by using Grammar Based Genetic Programming systems such as Grammatical Evolution  Develop AI agent for the Mario AI Benchmark following Grammatical Evolution

14.  Specify syntax of possible solutions through a context free grammar  Variable length integer strings are evolved following a genetic algorithm and these then choose the production rules from the grammar until all symbols are mapped

15.  Open source software based on the Mario World  All information given in two matrices (21x21)  Data about geometry of level and enemies that populate it is given  Information about current state of game is also given (position, status, mode etc)  Mario has 6 effectors (up, down, left, right, jump and fire and jump)

16.  At every cycle a button needs to be pressed in order to move Mario – impacts execution of BT  Control Nodes ◦ Sequence – logical AND ◦ Selector – logical OR ◦ Filters - added to add loops  Leaf Nodes ◦ Conditions – use level info to perform checks ◦ Actions – Mario’s possible movements ◦ Sub-Trees – designed to solve specific problems (long jumps)

17.  The BT syntax was coded into the grammar ◦ 30 conditions ◦ 8 actions ◦ 19 sub-trees ◦ 4 filters  Evolution combines these as long as syntax is respected  Had to limit certain rule combinations through the grammar since some trees were impossible to read and too demanding to execute

18.  Trees can have variable length but follow an and-or structure  Trees consist of a root node and number of sub trees (behaviour blocks)  Each block consists of one or more conditions followed by a sequence of actions  Tree has a default sub tree that is selected if no conditions are satisfied  If no default then it could be the case that Mario does not move

19.  Each behaviour block is self contained and hence it allows for individuals to exchange these between them in order to produce different behaviours  Use two point crossover  Allow for sub tree swap operation also (internal crossover)

20.  Each individual evaluated on 18 levels  Elitism ensures % of population kept from one generation to the next  At the end of all runs, all best individuals were evaluated on 600 unseen maps

21.  Best BT generated comprised of four behaviour blocks

22.  BT sent to Mario AI competition and managed to score in fourth place, proving BTs validity in the field of AI (relatively high kill count)  First and third place used A* variants  Second place used a neural network

23.  Use of Grammar simplifies the task of encoding the syntax of the BTs  Remarkable reactive behaviour capabilities but does not excel at planning  Hybrid approach under construction to aid in path planning

24.  DEFCON starts with a set of hand crafted trees, encoding feasible behaviours for each of the game four parts  Separate genetic programs were run for each part, creating new predefined behaviours from the original set  Main difference is that the BTs used in the Mario agent were encoded into a context free grammar in order to reduce complexity and create the behaviour blocks  Mario BTs also have loops defined in structure