Evolving Behaviour Trees for different AI-

bots Nicholas Mifsud

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

Introduction

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)

Behaviour Trees

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

BTs for DEFCON

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

DEFCON

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

DEFCON BTs

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

Genetic Operators

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)

Fitness Functions

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%

Experiment

All results indicate the mean of each fitness function increased as more generations were produced

Higher percentage win rate

Results

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%

Results

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)

Conclusions and Future Work

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

BTs using Grammatical Evolution

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

Grammatical Evolution

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)

Mario AI Benchmark

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)

Behaviour Trees for Mario

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

Incorporation into GE

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

Incorporation into GE

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)

Genetic Operators

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

Experiments

Best BT generated comprised of four behaviour blocks

Results

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

Competition

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

Conclusion and Future Work

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

Comparison of BTs Approach