v0.19.0

QuantEcon v0.19.0

Diff since v0.18.0

New features

• DiscreteDP: state_values and action_values label fields (constructor keywords), with the value-keyed decode layer sigma_values, state_to_index, DDPPolicyFunction, DDPValueFunction, and the IndexMap utility.
• MarkovChain(ddp, sigma): the Markov chain controlled by the policy sigma.

Breaking changes

• DPSolveResult is now immutable, with new fields ddp, epsilon, max_iter, and k and two appended type parameters; the (ddp, ddpr)-taking methods are removed (compute_greedy! entirely) — migrate to the array-level forms, e.g. bellman_operator!(ddp, res.v, res.Tv, res.sigma).
• The solution methods and backward_induction now run in the promoted Bellman value type: mixed-precision models (e.g. Float32 rewards with a Float64 beta) return promoted value arrays and may select different (correct) policies at near-tied action values.

Merged pull requests:

• ENH: Make DPSolveResult immutable with solve metadata (#401) (@oyamad)
• ENH: Add state_values/action_values to DiscreteDP (#402) (@oyamad)
• FIX: Run the DDP solver iterations in the promoted Bellman value type (#404) (@oyamad)

Closed issues:

• DiscreteDP: Add state_values, action_values (#120)
• DPSolveResult: record solution metadata; decide on Tv and mc cleanups (#390)
• FIX: VFI and MPFI compute state-action values in a narrower type than the public Bellman operator (#403)