Symbol manipulation by internal simulation of perception and behaviour

In traditional cognitive science, cognition amounts to symbol manipulation (Newell and Simon, 1972). Symbol manipulation is the processing of symbolic descriptions to produce an output that benefits an objective. Moreover, embodied models of cognitive development concern situated robots that do not rely on symbolic descriptions or on their manipulation (see, e.g., (Schlesinger, 2003)). In this paper, we introduce an embodied model of cognitive development that does rely on symbol manipulation. It has been claimed that internalised interaction with the external environment constitutes symbol manipulation (Phaf and Wolters, 1997). A mechanism to internalise such interaction was suggested in (Hesslow, 2002) as part of the simulation hypothesis. The simulation hypothesis states that conscious thought is based on the ability to simulate perception and behaviour internally. Such internal simulation consists of imagining actions and their consequences without actually performing them. In (Ziemke et al., 2005) it was demonstrated that situated robots can perform simple maze-following tasks on the basis of internal simulation. In a study involving situated agents engaged in an active categorisation task (van Dartel et al., 2005), we have shown that situated robots with the ability to simulate perception and behavior internally can outperform robots that do not have this ability (van Dartel et al., 2004). Situated robots can thus exploit their ability to simulate perception and behaviour internally. As was stated above, this ability may also constitute symbol manipulation. Symbol manipulation is often associated with planning (Newell and Simon, 1972), a skill that is generally regarded as high-level cognition (Cooper, 2002). To plan ahead in time, one needs to be able to represent the current state of the task in symbols and extrapolate to future states by manipulation of these symbols. The Tower of London (ToL) task is a typical planning task (Shallice, 1982) and a standard neuropsychological test to assess frontal lobe damage (Kolb and Whishaw, 1983), which impairs planning performance (Baddeley, 1986). In this paper we investigate whether symbol manipulation tasks can be solved by robots with the ability to simulate perception and behaviour internally. In order to do so, we constructed the situated model of the Tower of London task called STOL and formulated the first research question: Does the ability to perform symbol manipulation by internal simulation enhance performance on the ToL task in SToL? If this appears to be the case, we will try to answer the second research question: How does the symbol manipulation by internal simulation in SToL enhance performance on the ToL task? The ToL task will be discussed in more detail in section 2. The STOL model is described in section 3. The experiments conducted with the STOL model and the results found are reported in section 4. An analysis of the internal simulation of the optimised robot will be conducted in section 5. Finally, the results are discussed and concluded upon in section 6.