void enterImmReward( int cur_state, int next_state, int obs,
REAL_VALUE value ) {
/* cur_state is ignored for a POMDP, and obs is ignored for an MDP */
assert( gCurImmRewardNode != NULL );
switch( gCurImmRewardNode->type ) {
case ir_value:
gCurImmRewardNode->rep.value = value;
break;
case ir_vector:
if( gProblemType == POMDP_problem_type )
gCurImmRewardNode->rep.vector[obs] = value;
else
gCurImmRewardNode->rep.vector[next_state] = value;
break;
case ir_matrix:
if( gProblemType == POMDP_problem_type )
addEntryToIMatrix( gCurIMatrix, next_state, obs, value );
else
addEntryToIMatrix( gCurIMatrix, cur_state, next_state, value );
break;
default:
fprintf( stderr, "** ERR ** Unreckognized IR_Type in enterImmReward().\n");
exit( -1 );
break;
} /* switch */
} /* enterImmReward */
void
computeRewards() {
int a, i, j, z, next_state, obs;
double sum, inner_sum;
/* For the some problems, where we may want to shift all the reward
values to remove negative rewards, it will help to maintain the
minimum reward. Because all unrepresented values are zero, this
is our starting point. */
gMinimumImmediateReward = 0.0;
/* Now do the expectation thing for action-state reward values */
for( a = 0; a < gNumActions; a++ )
for( i = 0; i < gNumStates; i++ ) {
sum = 0.0;
/* Note: 'j' is not a state. It is an index into an array */
for( j = P[a]->row_start[i];
j < P[a]->row_start[i] + P[a]->row_length[i];
j++ ) {
next_state = P[a]->col[j];
if( gProblemType == POMDP_problem_type ) {
inner_sum = 0.0;
/* Note: 'z' is not a state. It is an index into an array */
for( z = R[a]->row_start[next_state];
z < (R[a]->row_start[next_state] + R[a]->row_length[next_state]);
z++ ) {
obs = R[a]->col[z];
inner_sum += R[a]->mat_val[z]
* getImmediateReward( a, i, next_state, obs );
} /* for z */
} /* if POMDP */
else /* it is an MDP */
inner_sum = getImmediateReward( a, i, next_state, 0 );
sum += P[a]->mat_val[j] * inner_sum;
} /* for j */
/* Update the minimum reward we are maintaining. */
gMinimumImmediateReward
= (gMinimumImmediateReward) < (sum)
? (gMinimumImmediateReward)
: (sum);
addEntryToIMatrix( IQ, a, i, sum );
} /* for i */
} /* computeRewards */
void computeRewards() {
int a, i, j, z, next_state, obs;
double sum, inner_sum;
/* Now do the expectation thing for action-state reward values */
for( a = 0; a < gNumActions; a++ )
for( i = 0; i < gNumStates; i++ ) {
sum = 0.0;
/* Note: 'j' is not a state. It is an index into an array */
for( j = P[a]->row_start[i];
j < P[a]->row_start[i] + P[a]->row_length[i];
j++ ) {
next_state = P[a]->col[j];
if( gProblemType == POMDP_problem_type ) {
inner_sum = 0.0;
/* Note: 'z' is not a state. It is an index into an array */
for( z = R[a]->row_start[next_state];
z < (R[a]->row_start[next_state] + R[a]->row_length[next_state]);
z++ ) {
obs = R[a]->col[z];
inner_sum += R[a]->mat_val[z]
* getImmediateReward( a, i, next_state, obs );
} /* for z */
} /* if POMDP */
else /* it is an MDP */
inner_sum = getImmediateReward( a, i, next_state, 0 );
sum += P[a]->mat_val[j] * inner_sum;
} /* for j */
addEntryToIMatrix( IQ, a, i, sum );
} /* for i */
} /* computeRewards */
