Driving in the wrong direction? Modelling policy mixes for EV adoption

Replication code for an agent-based model (ABM) of the transition from internal combustion engine vehicles (ICEVs) to electric vehicles (EVs), calibrated on California, 2001–2023.

The model couples four submodules so that consumer adoption and firm innovation co-evolve:

• Discrete choice consumption: heterogeneous car users compare the lifecycle utility (quality, range, lifecycle costs, lifecycle emissions) of their own car against new and used alternatives, and choose via a logit model.
• Social imitation: users only consider EVs once the share of EV owners in their Watts–Strogatz network neighbourhood passes an idiosyncratic innovativeness threshold.
• Directed innovation: manufacturers search parallel NK landscapes for ICEVs and EVs, choosing research direction and product mix by expected profit across consumer segments.
• Used car market: a single consolidated dealer prices used cars off the most similar new car with age-based depreciation, and scraps cars below scrap value.

Policy experiments evaluate five market-based instruments (carbon price, electricity subsidy, new EV rebate, used EV rebate, and production subsidy) individually and in pairs, over 2024–2035, with a post-policy projection to 2050.

---

Installation

The project is managed with uv and requires Python 3.13+.

uv sync

This installs a CPU-only build of PyTorch (only sbi needs it, for the simulation-based-inference calibration) rather than the default CUDA build; see the comments in pyproject.toml.

Running

All scripts resolve config paths relative to the repository root and use absolute imports, so run them as modules from the root:

uv run python -m package.generating_data.calibration_gen

Each *_gen.py script writes a timestamped directory under results/ and most call their matching *_plot.py immediately afterwards. results/ is git-ignored, so a fresh clone reproduces outputs by re-running the generators.

To explore a single run interactively, use model_playground.ipynb: load a base_params dict from package/constants/, call generate_data(base_params), and inspect the returned controller.

---

Reproducing the paper

Reproduction is two steps. First run the generator for a figure, which leaves a timestamped folder under results/. Then paste that folder name into the RUNS dict at the top of package/paper_figures/build_figures.py and run the builder, which re-plots from the results folder, renumbers each PNG to its number in the paper, and recomputes the results tables:

uv run python -m package.paper_figures.build_figures --list          # what is ready, what is missing
uv run python -m package.paper_figures.build_figures --only 3,S6     # build a subset

--list is the quickest way to see which RUNS entries are still empty. The builder writes into the manuscript directory (docs/paper/), which is not part of this repository — without it, use --no-tex to produce the renumbered PNGs alone, or take each figure straight from its results folder. See package/paper_figures/README.md for the flags and the full figure manifest.

Runs use 64 Monte Carlo seeds and are parallelised over available cores with joblib; the policy experiments are the expensive ones. Every *.slurm file next to a generator is the cluster version of the same command.

Manuscript figures and tables

Paper output	Generate	Plot	Config
Fig. 2: calibration 2001–2023 (EV uptake and sales, prices, HHI, car age)	generating_data.calibration_gen	plotting_data.calibration_plot	base_params_calibration.json
Fig. 3, Table 2: single-instrument grid search (100 intensities per instrument); Table 2 lists the maximum intensity per instrument, which is the upper bound in the bounds file	analysis.vary_single_policy_gen	analysis.vary_single_policy_plot	base_params_vary_single_policy_gen.json, analysis/policy_bounds_vary_single_policy_gen.json
Table 3: minimum single-policy intensity reaching 95% uptake, and its outcomes	analysis.endogenous_policy_intensity_single_gen	analysis.endogenous_policy_intensity_single_plot	base_params_endogenous_policy_single_gen.json, analysis/policy_bounds_endog_single_gen.json
Fig. 4: policy pairs achieving 94–96% uptake	analysis.endogenous_policy_intensity_pair_gen	analysis.endogenous_policy_intensity_pair_plot	base_params_endogenous_policy_pair_gen.json, analysis/policy_bounds_vary_pair_policy_gen.json
Fig. 5: trajectories to 2050 after policy removal	analysis.low_policy_intensity_gen	analysis.low_policy_intensity_plot	reads pair- and single-analysis results folders (see note below)
Fig. 6: parameter distribution histograms	generating_data.single_experiment_gen	plotting_data.single_experiment_plot	inline base_params dict in the script
Fig. 7: EV uptake vs used car market capacity	generating_data.vary_single_param_gen	plotting_data.vary_single_param_plot	base_params_vary_single.json, vary_single_max_num_cars_prop.json
BAU reference outcomes	analysis.BAU_outcomes_gen	n/a	base_params_endogenous_policy_pair_gen.json

Figure 1 is a hand-drawn model diagram, not model output. build_figures.py also recomputes an emissions table (tab:emissions) that the working manuscript carries but the submitted version drops.

Supplementary figures

package/supplementary_runs/ reproduces the supplementary material against the current package/constants/base_params_calibration.json. Every script there is a thin wrapper around generation and plotting code that already exists elsewhere in package/; each file's docstring names the function it calls.

Figure	What it shows	Script	Config
S1	NPE posterior density for a_chi, b_chi	calibration.sbi_single_seed_gen, plotted by supplementary_runs.fig01_posterior_plot	base_params_NN.json
S5	Simulated vs real-world ICE/EV price and range	supplementary_runs.fig05_calibration_cars_gen	base_params_calibration.json
S6	10-panel local sensitivity of EV uptake	supplementary_runs.fig06_panels (resumable) or fig06_local_sensitivity_gen, then fig06_local_sensitivity_plot	base_params_vary_single.json, vary_single_*.json
S7, S8	Sobol first- and total-order indices, 6 outputs × 10 parameters	supplementary_runs.fig07_08_sobol_gen	base_params_SA.json, variable_parameters_dict_SA.json
S9, S10	BAU EV uptake, emissions and elasticities over grid decarbonisation × electricity price	supplementary_runs.fig09_10_bau_gen, then fig09_bau_timeseries_plot and fig10_bau_elasticity_plot	base_params_inputs_and_emissions.json, vary_sen_decarb.json, vary_sen_elec_price.json
S11–S14	EV uptake in 2035 over a behavioural parameter × a policy intensity	supplementary_runs.fig11_14_policy_grid_gen <11|12|13|14>	base_params_vary_policy_joint.json, vary_policy_*.json

Figures S2–S4 are external data and NK-landscape illustrations, not model output. package/supplementary_runs/README.md has the run counts, wall-clock estimates, memory sizing and the SLURM jobs, including how to resume a partial Figure S6 and how to reuse one BAU sweep across a pair of policy grids.

Ordering note

The policy stages are sequential: the single-instrument analysis produces the intensity bounds used by the pair analysis, and low_policy_intensity_gen consumes the output folders of both. Paste those folder names into the ENDOG_PAIR and ENDOG_SINGLE constants at the top of package/analysis/low_policy_intensity_gen.py, or pass them on the command line:

uv run python -m package.analysis.low_policy_intensity_gen \
    endog_pair_<timestamp> --single-policy endog_single_<timestamp>

Policy intensities are found by Bayesian optimisation over a Gaussian-process surrogate (skopt.gp_minimize), maximising expected improvement against the 95% uptake target with a 1% tolerance.

Note that single_experiment_gen.py defines its parameters as an inline dict rather than loading a JSON config; it is the most convenient place to read off the full parameter set of Appendix B in one piece.

Not tied to a paper output

generating_data/ also holds exploratory sweeps kept for reference: policy_sensitivity_gen, sen_vary_single_param_gen (and its _second_hand_cars variant), battery_cost_sen_gen, sweep_hhi_age_gen, ablation_gen, burn_in_ablation_gen, delta_carbon_price_gen, and the single-parameter sweeps a_chi_sweep_gen, b_chi_grid_search_gen, delta_sweep_gen, kappa_sweep_gen and seed_inputs_sweep_gen. analysis/policy_dominance.py compares policy pairs across result folders, and calibration/NN_multi_round_calibration_multi_gen.py is the earlier multi-round form of the SBI calibration that sbi_single_seed_gen.py replaced.

---

Repository layout

├── model_playground.ipynb              # Interactive single-run exploration
├── pyproject.toml / uv.lock            # Dependencies (uv)
├── docs/
│   ├── code_narrative.tex / .pdf       # Extended walkthrough of the model code
│   └── forward_looking/                # Note on the forward-looking expectations extension
└── package/
    ├── model/                          # Core agent-based model
    ├── analysis/                       # Policy experiments (paper Section 4)
    ├── generating_data/                # Calibration, sensitivity and sweep runners
    ├── plotting_data/                  # Figure scripts for the above
    ├── paper_figures/                  # Renumbers figures and tables into the manuscript
    ├── supplementary_runs/             # Reproduces the supplementary figures
    ├── calibration/                    # Empirical data loading and SBI calibration
    ├── calibration_data/               # California input data (see Data sources)
    ├── constants/                      # base_params_*.json and vary_*.json configs
    └── resources/                      # Run harness and I/O helpers

package/model/

File	Role
controller.py	Orchestrates the monthly update sequence, exogenous input paths, and policy application
socialNetworkUsers.py	Car users: choice set construction, logit choice, imitation threshold update
firm.py	A single manufacturer: product mix, pricing, NK innovation
firmManager.py	Firm population and market segmentation
nkModel_ICE.py, nkModel_EV.py	NK technology landscapes for each drivetrain
carModel.py	A car design offered for sale (attribute vector plus price)
personalCar.py	An owned vehicle and its accumulated state
secondHandMerchant.py	Used car pricing, stock limits and scrapping
VehicleUser.py	Base user attributes
centralizedIdGenerator.py	Unique IDs across the simulation

package/resources/

• run.py: generate_data() builds and steps a controller for one parameter set; load_in_controller() resumes a calibrated controller for the policy period, so the 2001–2023 burn-in and calibration are computed once and reused across policy scenarios.
• utility.py: object save/load, run naming, directory creation, worker counts.

Configuration

Each experiment loads a base_params_*.json from package/constants/. Runs are divided into phases by timestep count: duration_burn_in (180, ICEV-only), duration_calibration (276, 2001–2023) and duration_future (144, the 2024–2035 policy period, extended to 2050 for the stability analysis, and 0 for the calibration and sensitivity configs that stop in 2023). All configurations use seed_repetitions 64.

Policies live under parameters_policies, with States switching each instrument on or off and Values giving its intensity. The vary_*.json files describe parameter sweeps consumed by the sensitivity runners.

---

Extensions present in the code but not used in the paper

The model contains three optional mechanisms that are disabled by default and switched off in every configuration shipped in this repository. They produce no result reported in the paper, and are retained as a starting point for follow-up work on command-and-control policy and policy anticipation. Each is guarded, so omitting its parameter reproduces the published behaviour exactly.

ICE bans (command-and-control)

The paper deliberately restricts itself to market-based instruments and excludes command-and-control policy. Three independent ban levers are nevertheless implemented, each taking effect a given number of months after the burn-in ends (the same convention as ev_production_start_time), and each defaulting to None, meaning "never":

Parameter	Effect
ICE_research_ban_time	Firms may no longer research or improve ICE technology. Existing ICE designs stay in firm memory and remain sellable.
ICE_sales_ban_time	Firms may no longer offer new ICE cars. Already-sold ICEVs remain drivable and resellable on the used market.
ICE_driving_ban_time, ICE_driving_ban_penalty	A per-unit cost shock added to effective ICE fuel cost from the ban date onward, reaching second-hand ICEVs too.

They are additive, so research-only, research+sales, and research+sales+driving scenarios can all be configured. The driving ban is modelled as a cost shock rather than a hard prohibition, which means it flows through the existing utility-driven choice mechanism and needs no new decision rule; the sales and research bans are hard constraints on firms' choice sets. All three validate that EV research or production has already begun before the ban bites.

Forward-looking expectations

forward_looking_expectations (default False) switches agents away from the naive expectations used in the paper, under which the current energy price and emissions intensity are assumed to persist indefinitely and the discounted sum collapses to the closed-form geometric series of Appendix A. With the flag on, agents instead discount the actual known future path of fuel costs, electricity prices and grid carbon intensity.

controller.compute_discounted_indices() precomputes present-value indices by backward recursion once per run, so the switch costs nothing per utility evaluation. Two subtleties are handled there: the carbon price is extended past the simulated horizon using its own schedule (so a temporary tax is not anticipated as a permanent one at its peak rate), while base energy prices and the decarbonisation trend are extended by holding their final value. With the flag off the indices are still computed but never read, so behaviour is unchanged.

This is the channel through which forward-looking agents would anticipate an announced ban or a scheduled carbon price ahead of its arrival, over a horizon implied by the model's own discount and depreciation rates rather than a separate anticipation parameter. docs/forward_looking/ works through the derivation.

Carbon price ramp and research subsidy

calculate_growth() supports flat, linear, quadratic and exponential carbon price paths via Carbon_price_state. Every configuration here uses flat, since the paper applies policies at full intensity from January 2024 and holds them constant, but the ramp is the hook for time-varying schemes such as the EU ETS. The ramp is coded to end at the close of the policy period rather than persist at its final value.

A sixth policy lever, Research_subsidy, is present in the policy state dictionary but is not among the five instruments analysed in the paper.

---

Data sources

package/calibration_data/ holds the California series used for calibration: vehicle population and EV sales (California Energy Commission), gasoline and residential electricity prices, grid emissions intensity, and CPI for conversion to 2020 US dollars. calibration_data_inputs.py assembles these into the pickled input object the model reads; calibration_data_outputs.py formats the observed targets used for indirect calibration.

Full parameter values, units and sources are tabulated in the paper's Appendix B.