Fig. 1. Task structure of Cannon Blast, a smartphone game to assess model-based planning.

a In this game, participants’ goal is to shoot as many diamonds as possible before their total number of shots (100 per block) runs out. To do so, they must aim a central cannon and then select which circular container to draw from. b Purple and pink balls dynamically bounce around each of the flanked containers which depict the probability of a pink or purple ball being released. For example, the left container displays 8 purple balls and releases a purple ball 80% of the time (a common transition) and displays 2 pink balls, giving a pink ball on 20% of trials (a rare transition). c The purple and pink balls have different values that dynamically change throughout the game. The value of the ball is defined as the probability of it being a ‘good ball’, i.e., one that remains intact after firing (rewarding trial), or a dud ball (non-rewarding trial) that explodes shortly after being fired, and therefore cannot reach the diamond. d We included 2 drifting reward probabilities (A, B) that quantitively differed on various metrics (see Supplementary Table 6). Participants were randomly assigned a reward drift set at each block leading to four distinct drift set combinations (A-A, A-B, B-A, B-B).