Poker-AI.org

int PureCfrMachine::walk_pure_cfr( const int position,
const BettingNode *cur_node,
const hand_t &hand,
rng_state_t &rng )
{
  int retval = 0;

  if( ( cur_node->get_child( ) == NULL )
      || cur_node->did_player_fold( position ) ) {
    /* Game over, calculate utility */
   
    retval = cur_node->evaluate( hand, position );
   
    return retval;
  }

  /* Grab some values that will be used often */
  int num_choices = cur_node->get_num_choices( );
  int8_t player = cur_node->get_player( );
  int8_t round = cur_node->get_round( );
  int64_t soln_idx = cur_node->get_soln_idx( );
  int bucket;
  if( ag.card_abs->can_precompute_buckets( ) ) {
    bucket = hand.precomputed_buckets[ player ][ round ];
  } else {
    bucket = ag.card_abs->get_bucket( ag.game, cur_node, hand.board_cards,
hand.hole_cards );
  }

  /* Get the positive regrets at this information set */
  uint64_t pos_regrets[ num_choices ];
  uint64_t sum_pos_regrets
    = regrets[ round ]->get_pos_values( bucket,
soln_idx,
num_choices,
pos_regrets );
  if( sum_pos_regrets == 0 ) {
    /* No positive regret, so assume a default uniform random current strategy */
    sum_pos_regrets = num_choices;
    for( int c = 0; c < num_choices; ++c ) {
      pos_regrets[ c ] = 1;
    }
  }

  /* Purify the current strategy so that we always take choice */
  uint64_t dart = genrand_int32( &rng ) % sum_pos_regrets;
  int choice;
  for( choice = 0; choice < num_choices; ++choice ) {
    if( dart < pos_regrets[ choice ] ) {
      break;
    }
    dart -= pos_regrets[ choice ];
  }
  assert( choice < num_choices );
  assert( pos_regrets[ choice ] > 0 );
 
  const BettingNode *child = cur_node->get_child( );

  if( player != position ) {
    /* Opponent's node. Recurse down the single choice. */

    for( int c = 0; c < choice; ++c ) {
      child = child->get_sibling( );
    }

    retval = walk_pure_cfr( position, child, hand, rng );

    /* Update the average strategy if we are keeping track of one */
    if( do_average ) {
      if( avg_strategy[ round ]->increment_entry( bucket, soln_idx, choice ) ) {
fprintf( stderr, "The average strategy has overflown :(\n" );
fprintf( stderr, "To fix this, you must set a bigger AVG_STRATEGY_TYPE "
"in constants.cpp and start again from scratch.\n" );
exit( 1 );
      }
    }
   
  } else {
    /* Current player's node. Recurse down all choices to get the value of each */

    int values[ num_choices ];
   
    for( int c = 0; c < num_choices; ++c ) {
      values[ c ] = walk_pure_cfr( position, child, hand, rng );
      child = child->get_sibling( );
    }

    /* We return the value that the sampled pure strategy attains */
    retval = values[ choice ];

    /* Update the regrets at the current node */
    regrets[ round ]->update_regret( bucket, soln_idx, num_choices,
values, retval );
  }
 
  return retval;
}