''Table of Contents''
#contents
* Overview [#xfb9bf26]
Here, we introduce how to implement a simple maze task with SkyAI.
The maze task has a discrete state and a discrete action, which will be implemented as a module of SkyAI.
As an reinforcement learning algorithm, Peng's Q(lambda)-learning is applied to the maze task; of course, we use predefined modules.
The following is the procedure:
+ Implement a maze task module.
+ Implement a random action module for testing the task module.
+ Implement a main function.
+ Compile.
+ Write an agent script for the random action test.
+ Write an agent script to apply Q(lambda)-learning.
The sample code works on a console; no extra libraries are required.
Let's start!
* Task Setup [#c0a22683]
The maze has the size W x H, consisting of the start (S), the goal (G), and the walls.
The robot cannot go through the walls.
Its objective is to move from the start to the goal in the shortest path.
This is an example of the maze environment:
# # # # # # # # # #
# # G #
# # # #
# S # # # #
# # # # # #
# # #
# #
# # # # # # # # # #
The ''state'' is a 1-dimensional discrete value where the (x, y) position is serialized.
The ''action'' is a discrete action consisting of {up,down,left,right}.
The ''reward'' is given +1 when the robot arrives at the goal, and -0.01 for each action.
Each episode starts with locating the robot at the start, and ends when the robot reaches the goal or the amount of steps becomes greater than a threshold (1000).
* Maze Task Module [#g1d7df27]
Please refer to [[../Tutorial - Making Module]].
+ Make a C++ source file named maze.cpp using a template materials/templates/apps/main_tmpl.cpp contained in the SkyAI directory.
-- You can modify the file information (file name, brief, author, date, copyright, license info, etc.)
-- Replace every NAME_SPACE by loco_rabbits.
-- Write the following code inside the namespace loco_rabbits.
+ Make a configure class using the template TXxConfigurations written in [[../Tutorial - Making Module]].
-- Replace every TXxConfigurations by TMazeTaskConfigurations.
-- Remove the TestC parameter and add the following parameters:
#codeh(cpp){{
int NumEpisodes; // number of episodes
int MaxSteps; // number of max action steps per episode
int StartX, StartY; // start position
double GoalReward; // goal reward
double StepCost; // cost for each action step
int SleepUTime; // duration for display
std::vector<std::vector<int> > Map; // Map[y][x], 0:free space, 1:wall, 2:goal, every element should have the same size
}}
-- Initialize them at the constructor as:
#codeh(cpp){{
TMazeTaskConfigurations (var_space::TVariableMap &mmap) :
NumEpisodes (1000),
MaxSteps (1000),
StartX (1),
StartY (1),
GoalReward (1.0),
StepCost (-0.01),
SleepUTime (1000)
{
Register(mmap);
}
}}
-- In the member function Register, insert them:
#codeh(cpp){{
ADD( NumEpisodes );
ADD( StartX );
ADD( StartY );
ADD( GoalReward );
ADD( StepCost );
ADD( SleepUTime );
ADD( Map );
}}
-- Add lora/variable_space_impl.h in the include list.
#codeh(cpp){{
#include <lora/variable_space_impl.h> // to store std::vector<TIntVector>
}}
-- You can add your own parameters such as a noise.
+ Make the base of the module using the template MXxModule written in [[../Tutorial - Making Module]].
-- Simple template is OK.
-- Replace every MXxModule by MMazeEnvModule.
-- Replace every MParentModule by TModuleInterface.
-- Replace TXxConfigurations by TMazeTaskConfigurations.
-- Remove the definition of mem_ (TXxMemory mem_;).
#codeh(cpp){{
//===========================================================================================
//!\brief Maze task (environment+task) module
class MMazeTaskModule
: public TModuleInterface
//===========================================================================================
{
public:
typedef TModuleInterface TParent;
typedef MMazeTaskModule TThis;
SKYAI_MODULE_NAMES(MMazeTaskModule)
MMazeTaskModule (const std::string &v_instance_name)
: TParent (v_instance_name),
conf_ (TParent::param_box_config_map())
{
}
protected:
TMazeTaskConfigurations conf_;
}; // end of MMazeTaskModule
//-------------------------------------------------------------------------------------------
}}
+ Add following ports into MMazeEnvModule.
-- (port type), (port name), (return type), (parameter list), (purpose)
-- slot, slot_start, void, (void), called at the beginning of the execution.
-- slot, slot_execute_action, void, (const TInt &a), called by an RL agent module to execute action.
-- signal, signal_initialization, void (void), emit when the module is initialized.
-- signal, signal_start_of_episode, void (void), emit when each episode starts.
-- signal, signal_finish_episode, void (void), emit when the end-of-episode condition is satisfied.
-- signal, signal_end_of_episode, void (void), emit when each episode is terminated.
-- signal, signal_start_of_step, void (void), emit at the start of each step.
-- signal, signal_end_of_step, void (void), emit at the end of each step.
-- signal, signal_reward, void (const TSingleReward &), emit when a reward is given.
-- out, out_state_set_size, const TInt&, (void), output the number of elements in the state set.
-- out, out_action_set_size, const TInt&, (void), output the number of elements in the action set.
-- out, out_state, const TInt&, (void), output the current state (x,y are serialized).
-- out, out_time, const TReal&, (void), output the current time.
-- Note: some signal ports will not be used, but, defined for later use.
-- In order to add the ports, follow the steps:
++ Add declarations:
#codeh(cpp){{
MAKE_SLOT_PORT(slot_start, void, (void), (), TThis);
MAKE_SLOT_PORT(slot_execute_action, void, (const TInt &a), (a), TThis);
MAKE_SIGNAL_PORT(signal_initialization, void (void), TThis);
MAKE_SIGNAL_PORT(signal_start_of_episode, void (void), TThis);
MAKE_SIGNAL_PORT(signal_finish_episode, void (void), TThis);
MAKE_SIGNAL_PORT(signal_end_of_episode, void (void), TThis);
MAKE_SIGNAL_PORT(signal_start_of_step, void (void), TThis);
MAKE_SIGNAL_PORT(signal_end_of_step, void (void), TThis);
MAKE_SIGNAL_PORT(signal_reward, void (const TSingleReward &), TThis);
MAKE_OUT_PORT(out_state_set_size, const TInt&, (void), (), TThis);
MAKE_OUT_PORT(out_action_set_size, const TInt&, (void), (), TThis);
MAKE_OUT_PORT(out_state, const TInt&, (void), (), TThis);
MAKE_OUT_PORT(out_time, const TReal&, (void), (), TThis);
}}
++ Add initializers at the constructor:
#codeh(cpp){{
MMazeTaskModule (const std::string &v_instance_name)
: ...
slot_start (*this),
slot_execute_action (*this),
signal_initialization (*this),
signal_start_of_episode (*this),
signal_finish_episode (*this),
signal_end_of_episode (*this),
signal_start_of_step (*this),
signal_end_of_step (*this),
signal_reward (*this),
out_state_set_size (*this),
out_action_set_size (*this),
out_state (*this),
out_time (*this)
}}
++ Add register functions at the constructor:
#codeh(cpp){{
add_slot_port (slot_start );
add_slot_port (slot_execute_action );
add_signal_port (signal_initialization );
add_signal_port (signal_start_of_episode );
add_signal_port (signal_finish_episode );
add_signal_port (signal_end_of_episode );
add_signal_port (signal_start_of_step );
add_signal_port (signal_end_of_step );
add_signal_port (signal_reward );
add_out_port (out_state_set_size );
add_out_port (out_action_set_size );
add_out_port (out_state );
add_out_port (out_time );
}}
+ Next, we implement the slot port callbacks and the output functions. This procedure is slightly complicated; follow one by one.
++ Add member variables at the protected section.
#codeh(cpp){{
mutable int state_set_size_;
const int action_set_size_;
int current_action_;
int pos_x_, pos_y_;
mutable int tmp_state_;
TReal current_time_;
TInt num_episode_;
}}
++ Add their initializers:
#codeh(cpp){{
state_set_size_ (0),
action_set_size_ (4),
current_action_ (0),
}}
++ Implement slot_start_exec. This is a long code, so, write the declaration at the protected section:
#codeh(cpp){{
virtual void slot_start_exec (void);
}}
Then, define it outside the class:
#codeh(cpp){{
/*virtual*/void MMazeTaskModule::slot_start_exec (void)
{
init_environment();
signal_initialization.ExecAll();
for(num_episode_=0; num_episode_<conf_.NumEpisodes; ++num_episode_)
{
init_environment();
signal_start_of_episode.ExecAll();
bool running(true);
while(running)
{
signal_start_of_step.ExecAll();
running= step_environment();
show_environment();
usleep(conf_.SleepUTime);
if(current_time_>=conf_.MaxSteps)
{
signal_finish_episode.ExecAll();
running= false;
}
signal_end_of_step.ExecAll();
}
signal_end_of_episode.ExecAll();
}
}
}}
where we used the three member functions. These are declared at the protected section:
#codeh(cpp){{
void init_environment (void);
bool step_environment (void);
void show_environment (void);
}}
and, defined outside the class:
#codeh(cpp){{
void MMazeTaskModule::init_environment (void)
{
pos_x_= conf_.StartX;
pos_y_= conf_.StartY;
current_time_= 0.0l;
}
}}
#codeh(cpp){{
bool MMazeTaskModule::step_environment (void)
{
int next_x(pos_x_), next_y(pos_y_);
switch(current_action_)
{
case 0: ++next_x; break; // right
case 1: --next_y; break; // up
case 2: --next_x; break; // left
case 3: ++next_y; break; // down
default: LERROR("invalid action:"<<current_action_);
}
++current_time_;
signal_reward.ExecAll(conf_.StepCost);
switch(conf_.Map[next_y][next_x])
{
case 0: // free space
pos_x_=next_x;
pos_y_=next_y;
break;
case 1: // wall
break;
case 2: // goal
pos_x_=next_x;
pos_y_=next_y;
signal_reward.ExecAll(conf_.GoalReward);
signal_finish_episode.ExecAll();
return false;
default: LERROR("invalid map element: "<<conf_.Map[next_y][next_x]);
}
return true;
}
}}
#codeh(cpp){{
void MMazeTaskModule::show_environment (void)
{
int x(0),y(0);
std::cout<<"("<<pos_x_<<","<<pos_y_<<") "<<current_time_<<"/"<<num_episode_<<std::endl;
for(std::vector<std::vector<int> >::const_iterator yitr(conf_.Map.begin()),ylast(conf_.Map.end());yitr!=ylast;++yitr,++y)
{
x=0;
for(std::vector<int>::const_iterator xitr(yitr->begin()),xlast(yitr->end());xitr!=xlast;++xitr,++x)
{
std::cout<<" ";
if(x==pos_x_ && y==pos_y_)
std::cout<<"R";
else if(x==conf_.StartX && y==conf_.StartY)
std::cout<<"S";
else
switch(*xitr)
{
case 0: std::cout<<" "; break;
case 1: std::cout<<"#"; break;
case 2: std::cout<<"G"; break;
default: std::cout<<"?"; break;
}
}
std::cout<<" "<<std::endl;
}
std::cout<<std::endl;
}
}}
++ Implement the other slot port callbacks and output functions. These are short code, so, you can write inside the class at the protected section.
#codeh(cpp){{
virtual void slot_execute_action_exec (const TInt &a)
{
current_action_= a;
}
virtual const TInt& out_state_set_size_get (void) const
{
state_set_size_= conf_.Map[0].size() * conf_.Map.size();
return state_set_size_;
}
virtual const TInt& out_action_set_size_get (void) const
{
return action_set_size_;
}
virtual const TInt& out_state_get (void) const
{
return tmp_state_=serialize(pos_x_,pos_y_);
}
virtual const TReal& out_time_get (void) const
{
return current_time_;
}
}}
where serialize is a protected member function defined as follows:
#codeh(cpp){{
int serialize (int x, int y) const
{
return y * conf_.Map[0].size() + x;
}
}}
+ Add a Start() public member function that calls slot_start:
#codeh(cpp){{
void Start()
{
slot_start.Exec();
}
}}
+ Finally, use SKYAI_ADD_MODULE macro to register the module on SkyAI:
#codeh(cpp){{
SKYAI_ADD_MODULE(MMazeTaskModule)
}}
This should be written outside the class and inside the namespace loco_rabbits.
That's it.
* Random Action Module [#x0a9632f]
Next, in order to test the MMazeTaskModule module, we make a module named MRandomActionModule that emits a random action at each step.
MRandomActionModule has two ports:
- (port type), (port name), (return type), (parameter list), (purpose)
- slot, slot_step, void, (void), called at each step where a random action is emitted through the signal_action port.
- signal, signal_action, void (const TInt &), emit at each step.
Thus, its implementation is very simple:
#codeh(cpp){{
//===========================================================================================
//!\brief Random action module
class MRandomActionModule
: public TModuleInterface
//===========================================================================================
{
public:
typedef TModuleInterface TParent;
typedef MRandomActionModule TThis;
SKYAI_MODULE_NAMES(MRandomActionModule)
MRandomActionModule (const std::string &v_instance_name)
: TParent (v_instance_name),
slot_step (*this),
signal_action (*this)
{
add_slot_port (slot_step );
add_signal_port (signal_action);
}
protected:
MAKE_SLOT_PORT(slot_step, void, (void), (), TThis);
MAKE_SIGNAL_PORT(signal_action, void (const TInt &), TThis);
virtual void slot_step_exec (void)
{
signal_action.ExecAll(rand() % 4);
}
}; // end of MRandomActionModule
//-------------------------------------------------------------------------------------------
}}
Then, use SKYAI_ADD_MODULE macro to register the module on SkyAI:
#codeh(cpp){{
SKYAI_ADD_MODULE(MRandomActionModule)
}}
* Main Function [#ddf2c0fe]
Refer to [[../Tutorial - Making Executable]].
Our main function is as follows:
#codeh(cpp){{
using namespace std;
using namespace loco_rabbits;
int main(int argc, char**argv)
{
TOptionParser option(argc,argv);
TAgent agent;
if (!ParseCmdLineOption (agent, option)) return 0;
MMazeTaskModule *p_maze_task = dynamic_cast<MMazeTaskModule*>(agent.SearchModule("maze_task"));
if(p_maze_task==NULL) {LERROR("module `maze_task' is not defined as an instance of MMazeTaskModule"); return 1;}
agent.SaveToFile (agent.GetDataFileName("before.agent"),"before-");
p_maze_task->Start();
agent.SaveToFile (agent.GetDataFileName("after.agent"),"after-");
return 0;
}
}}
This main function consists of the following parts:
+ Create an instance of the TAgent class.
+ Parse the command line option and load an agent script.
+ Get a module named maze_task which is an instance of MMazeTaskModule.
+ Save the agent status into a file named before.agent.
+ Execute the maze_task's Start function.
+ Save the agent status into a file named after.agent.
* Compile [#ga57fd4c]
First, write a makefile as follows:
#codeh(makefile){{
BASE_REL_DIR:=../..
include $(BASE_REL_DIR)/Makefile_preconf
EXEC := maze.out
OBJS := maze.o
USING_SKYAI_ODE:=true
MAKING_SKYAI:=true
include $(BASE_REL_DIR)/Makefile_body
}}
- BASE_REL_DIR : relative path to the base directory of the SkyAI.
Then, execute the make command:
#codeh(sh){{
make
}}
An executable named maze.out is generated?
* Agent Script for Random Action Test [#vdba6662]
Please refer to [[../Tutorial - Writing Agent Script]].
Now, let's test MMazeTaskModule using MRandomActionModule.
+ Create a blank file named random_act.agent and open it.
+ Instantiate each module; the MMazeTaskModule's instance should have the name maze_task:
#codeh(cpp){{
module MMazeTaskModule maze_task
module MRandomActionModule rand_action
}}
+ Connect the following port pairs:
-- maze_task.signal_start_of_step --> rand_action.slot_step
-- rand_action.signal_action --> maze_task.slot_execute_action
#codeh(cpp){{
connect maze_task.signal_start_of_step , rand_action.slot_step
connect rand_action.signal_action , maze_task.slot_execute_action
}}
+ Assign the maze information to the configuration parameters of maze_task:
#codeh(cpp){{
maze_task.config={
Map={
[]= (1,1,1,1,1,1,1,1,1,1)
[]= (1,0,0,0,1,0,0,0,2,1)
[]= (1,0,1,0,1,0,0,0,0,1)
[]= (1,0,1,0,1,1,0,0,0,1)
[]= (1,0,1,0,0,1,0,1,1,1)
[]= (1,0,0,0,0,1,0,0,0,1)
[]= (1,0,0,0,0,0,0,0,0,1)
[]= (1,1,1,1,1,1,1,1,1,1)
}
StartX= 1
StartY= 3
}
}}
That's it. Let's test!
Launch the executable as follows:
#codeh(sh){{
./maze.out -agent random_act
}}
You will see a maze as follows where the robot (R) moves randomly.
(1,5) 77/4
# # # # # # # # # #
# # G #
# # # #
# S # # # #
# # # # # #
# R # #
# #
# # # # # # # # # #
* Agent Script for Q(lambda)-learning [#ef68204c]
Please refer to [[../Tutorial - Writing Agent Script]].
If you can make sure that MMazeTaskModule works correctly, then, let's apply a Q-learning module.
+ Create a blank file named ql.agent and open it.
+ Include ql_dsda where a composite Q-learning module is defined:
#codeh(cpp){{
include_once "ql_dsda"
}}
+ Instantiate the modules; the MMazeTaskModule's instance should have the name maze_task:
#codeh(cpp){{
module MMazeTaskModule maze_task
module MTDDiscStateAct behavior
}}
+ Connect the port pairs:
#codeh(cpp){{
/// initialization process:
connect maze_task.signal_initialization , behavior.slot_initialize
/// start of episode process:
connect maze_task.signal_start_of_episode , behavior.slot_start_episode
/// learning signals:
connect behavior.signal_execute_action , maze_task.slot_execute_action
connect maze_task.signal_end_of_step , behavior.slot_finish_action
connect maze_task.signal_reward , behavior.slot_add_to_reward
connect maze_task.signal_finish_episode , behavior.slot_finish_episode_immediately
/// I/O:
connect maze_task.out_action_set_size , behavior.in_action_set_size
connect maze_task.out_state_set_size , behavior.in_state_set_size
connect maze_task.out_state , behavior.in_state
connect maze_task.out_time , behavior.in_cont_time
}}
+ Assign the maze information to the configuration parameters of maze_task:
#codeh(cpp){{
maze_task.config={
Map={
[]= (1,1,1,1,1,1,1,1,1,1)
[]= (1,0,0,0,1,0,0,0,2,1)
[]= (1,0,1,0,1,0,0,0,0,1)
[]= (1,0,1,0,1,1,0,0,0,1)
[]= (1,0,1,0,0,1,0,1,1,1)
[]= (1,0,0,0,0,1,0,0,0,1)
[]= (1,0,0,0,0,0,0,0,0,1)
[]= (1,1,1,1,1,1,1,1,1,1)
}
StartX= 1
StartY= 3
}
}}
+ Assign the learning configuration to the parameters of behavior:
#codeh(cpp){{
behavior.config={
UsingEligibilityTrace = true
UsingReplacingTrace = true
Lambda = 0.9
GradientMax = 1.0e+100
ActionSelection = "asBoltzman"
PolicyImprovement = "piExpReduction"
Tau = 1
TauDecreasingFactor = 0.05
TraceMax = 1.0
Gamma = 0.9
Alpha = 0.3
AlphaDecreasingFactor = 0.002
AlphaMin = 0.05
}
}}
Launch the executable as follows:
#codeh(sh){{
./maze.out -path ../../benchmarks/cmn -agent ql -outdir result/rl1
}}
where ../../benchmarks/cmn is a relative path of the benchmarks/cmn directory; modify it for your environment.
After several tens of episodes, the policy will converge to a path:
#block
(1,4) 1/520
# # # # # # # # # #
# # G #
# # # #
# S # # # #
# R # # # # #
# # #
# #
# # # # # # # # # #
#block(next)
(3,6) 5/520
# # # # # # # # # #
# # G #
# # # #
# S # # # #
# # # # # #
# # #
# R #
# # # # # # # # # #
#block(next)
(6,6) 8/520
# # # # # # # # # #
# # G #
# # # #
# S # # # #
# # # # # #
# # #
# R #
# # # # # # # # # #
#block(end)
#block
(7,3) 12/520
# # # # # # # # # #
# # G #
# # # #
# S # # # R #
# # # # # #
# # #
# #
# # # # # # # # # #
#block(next)
(8,1) 15/520
# # # # # # # # # #
# # R #
# # # #
# S # # # #
# # # # # #
# # #
# #
# # # # # # # # # #
#block(next)
#block(end)
In order to store the learning logs, make a directory result/rl1 which is specified with -outdir option.
Plotting log-eps-ret.dat, you will obtain a learning curve:
#ref(./out-maze.png,zoom,center,600x0)
CENTER:''Example of a learning curve.''
* Exercise [#h5477405]
+ Execute several runs (e.g. 10 runs) and plot their mean learning-curve; plot its deviation over the mean.
+ Test another parameters, algorithms, and maze kinds.
+ Extend the maze; e.g. include a trap, wind, etc.
+ Add noise at each step. The noise parameter should be contained in the configuration box.
+ Change the agent script to log the state transition in each step, and visualize the paths in an learning.
-- Check the logger modules MSimpleDataLogger1_T, MSimpleDataLogger2_T1_T2, MUniversalDataLogger where T, T1, and T2 denote a type.
![[SkyAI]](image/logoxs.png "[SkyAI]")
