In a set of models incorporating a varying amount of immunological complexity, we investigate processes by which T lymphocytes ‘learn’ to discriminate self from nonself. The models all incorporate the following immunological data: (1) T lymphocytes produce their own growth factor (IL2), (2) T lymphocyte effectors proliferate in response to IL2 and antigen, and (3) T lymphocyte effectors become memory cells whenever antigenic restimulation is poor. The first two facts are sufficient to generate a proliferation threshold: the (self-reinforeing) proliferation process only starts when LL2 concentrations are sufficiently high. We first analyse a fairly simple model i.e. an immune system in which cytotoxic effectors (CTL) and helper T cells (HTL, the cells that produce IL2) are assumed to be identical. This simple model accounts for self–nonself discrimination in the absence of any down-regulatory interactions (i.e. suppressions). Clones with high affinity to a self antigen remain below the proliferation threshold, whereas T lymphocytes with low self affinity accumulate memory cells and cross the proliferation threshold whenever a (high-affinity) foreign antigen enters the immune system. Secondly, we analyse this tolerance process in various more complex models that incorporate CTL and HTL as separate populations. Such an extension of the model adds interesting new features, because the populations now compete for andtigen and IL2. Competition for IL2 markedly influences tolerance: enlarged CTL populations (due, for example, to memory accumulation) raise the proliferation threshold by intensive IL2 absorption. We investigate here how such complex models can regain the tolerance behaviour of the previous (simple) model. This is achieved by incorporating the additional complexity of antigen presentation and a lymphoid factor that regulates the expression of the IL2 receptor. It thus appaears that results obtained in a simple model, which is attractive from a theoretical point of view, do require modification before their extrapolations to more complex systems is justified. Self-nonself discrimination develops in the models as a result of nonlinearities in the T lymphocyte proliferation and activation process. The parameter setting of the models was fixed; since the models incorporate the same immunological processes, the parameters should also be identical. The models were analysed by numerical methods: numerical integration and the numerical generatrion of zero-isoclines. We think that, by combining these methods, we can investigate a fairly large set of complex models concomitantly. The analysis of such a diverse set of models is markedly facilitated when the complex models are projected into 2-D or 3-D state spaces by quasi-steady-state assumptions: zero-isoclines in such reduced state spaces indicate the similarities and discrepancies between the various models.