Hoppers tend to be worth only half of the corresponding rider. (Compare the classical values of Rook and Cannon in Xiangqi, showing what it does to require a hop for the captures, and then imagine what would happen if you handicap the non-captures the same way...) The AI guestimates the value basically from counting of moves the piece would have on a 25% filled board, assuming a linear increase of the enemy density with rank, counting captures double, and assuming somewhat above-average placement of the piece. And then applies a modest non-linear correction to account for synergy. The latter would not reorder the values, though. So if it guestimates a non-divergent piece to have a higher value, it means that that piece will in general have more moves under conditions likely to occur in the decisive game phase.
Hoppers tend to be worth only half of the corresponding rider. (Compare the classical values of Rook and Cannon in Xiangqi, showing what it does to require a hop for the captures, and then imagine what would happen if you handicap the non-captures the same way...) The AI guestimates the value basically from counting of moves the piece would have on a 25% filled board, assuming a linear increase of the enemy density with rank, counting captures double, and assuming somewhat above-average placement of the piece. And then applies a modest non-linear correction to account for synergy. The latter would not reorder the values, though. So if it guestimates a non-divergent piece to have a higher value, it means that that piece will in general have more moves under conditions likely to occur in the decisive game phase.