Animal Decision-Making: Sequential Versus Simultaneous Choice

We are building a theory of decision-making, and to do this we investigate how birds take decisions. In behavioural science a decision is not assumed to result from thoughtful deliberation but refers to what organisms do when several actions are possible. In humans, introspection can create the perception (often inaccurate) that one's own decisions are driven by evaluation of each alternative, and as a consequence people often assume that animals also choose by evaluating alternatives. If this were true, choosing would take information-processing effort and time: more options, more time. For instance, if a lion sees a zebra, it may start a chase, but if it sees both a zebra and a wildebeest, it would evaluate their relative merits. If it chooses the zebra, it would have taken longer to start the chase. The paradox of choice, for instance, states that more options make choice more difficult. We found that the opposite is true in starlings: they take longer to take a lonely alternative than when they take the same alternative in a choice situation. In our experiments the times taken to take each option when faced alone predict very accurately how long it takes when it takes it out of a choice, and the shortening of time in choices results from the way the model works. Further, the time to accept each option when it is met alone depends not only on its absolute properties, but also on the benefit it gives relative to the context. To deal with all these findings, we used ideas originating in biology, economics and psychology to propose the Sequential Choice Model or SCM for short. SCM postulates that the mechanisms used by birds to choose between options are the same they use when facing each option alone. SCM incorporates the idea that these mechanisms evolved as adaptations to environments in which meeting different options simultaneously is rare, but meeting them sequentially is common. Thus, there are no special adaptations for simultaneous choices, but the time to chase each alternative is precisely tuned to exploit the benefits it gives compared with the opportunities in the whole environment. The SCM very effectively explained and predicted (post-hoc) the results in our original experiments, but the real value of a theoretical model is when it works for situations different from those that led to its inception. We propose to test SCM in choice problems that have never been studied from this perspective and to see if we still observe the same predictive precision including the shortening of decision times in choices. We'll use experiments that require cognition that might be expected to be time consuming. In one of them, a blue light is shown for a time lasting between 0 and 30 s, and after that either a red or a green light shows (In separate trials). If it is red, after the bird pecks the key it gets food after waiting 15 s, but if it is green the waiting time is 30 s minus the time the blue light had been on. Thus, if blue lasted 10 s, then green's waiting is 20 s, but if blue lasted 25 s, then green's waiting is 5 s. We measure how long the starling takes to peck in both red and green no-choice trials. On other (choice) trials, after the blue light goes off both red and green show, and we look at which one the bird chooses and how long it takes. To minimise waiting for food, birds should choose red if blue lasted less than 15 s and green if it lasted longer, but they don't do exactly this. SCM predicts what they will do using the times to peck red or green in no-choice trials, and it also predicts how long it will take to peck either: it should take less in choice than in no-choice trials. Since choice involves consulting the memory for the duration of blue one might expect choice to take extra time, but SCM predicts the opposite. If the SCM predictions are met, this would be evidence that it applies to very different situations from those in which it originated, and hence that it is a very valuable model.

We have recently proposed a model of animal choice (Sequential Choice Model, SCM) based on our previous work on starlings. I now intend to extend the model and to test its generality applying it to a wider range of experimental situations. SCM's main feature is that it predicts behaviour in choice situations using data from no-choice encounters with each alternative. Its premises are: (1) When an animal faces a single option, it doesn't take it immediately (the 'latency'). Each alternative faced on its own elicits a specific probability density function of latencies. Latencies are not reaction times: they exceed RTs duration by an order of magnitude and they have different properties. (2) Latencies to take single options are decreasing functions of the improvement in state-dependent fitness (or utility) that the decision maker expects from that option relative to the context. This feature connects the model to optimality theory. (3) Expectations about each option depend on both the subject's state and the average properties of the environment prevalent during learning. (4) When more than one option is met simultaneously, each elicits a sample from its original distribution of latencies. The shortest sample is expressed as a choice. There is no comparative evaluation at choice time: each option elicits a candidate latency just as in sequential encounters. This cross-censorship between latency distributions means that latencies for each option are shorter when picked out of a choice than when picked in the absence of alternatives. This is the opposite of how reaction times behave. The SCM was proposed for a system with pairs of options, where its predictive performance was extremely successful. To investigate its generality, we will now test it using the same species in a wide variety of choice paradigms, including multi-alternative choice, the time-left procedure, risky choice and comparative valuation scenarios.